Risk factors affecting pedestrian behaviour: risk assessment in a virtual environment
Abstract
A promoção de modos de transporte suaves, dos quais fazem parte andar a pé ou de bicicleta, ultimamente tem sido impulsionada devido às vantagens sociais e ambientais que estes possuem. Por outro lado, o aumento dos volumes de tráfego motorizado verificado nos últimos anos tem-se traduzido num aumento da exposição ao risco de acidente para todos os utilizadores da rede viária. No que se refere aos peões, esse crescimento leva a que ocorram mais interações entre estes utilizadores da rede viária e os veículos motorizados. O número de mortes de peões que ocorrem nas estradas está longe de ser nulo, sendo que uma parte considerável delas acontece em passagens para peões. Procurar prevenir a ocorrência de acidentes que possam ter graves consequências, promovendo, desta forma, maior conforto e segurança para todos os utilizadores da rede viária deve ser uma prioridade, tendo ciente que o paradigma da mobilidade está a mudar. Aproveitando as mais recentes tecnologias para a aquisição de dados sobre o comportamento de peões, como a análise automatizada de vídeo e o uso de um simulador onde num ambiente virtual se consegue estudar o comportamento dos peões quando executam o atravessamento da estrada sem que enfrentem um perigo real para a sua integridade física, o principal objetivo deste projeto de doutoramento foi a identificação e análise de fatores com influência no risco para peões quando atravessam a faixa de rodagem relacionados com as características das infraestruturas rodoviária e pedonal, as características do tráfego motorizado e pedonal e com o ruído emitido pelos veículos, dando particular relevância à tomada de decisão de atravessamento e à sua interação com os veículos. De uma forma geral, os resultados mostram a influência direta ou indireta dos diversos fatores abordados na tomada de decisão dos peões e na interação entre eles e os veículos que deles se aproximam aquando do atravessamento da faixa de rodagem. No entanto, os resultados que mais se evidenciam levam a concluir que a velocidade e a cinemática dos veículos em aproximação à passagem para peões são fatores com elevada importância para a segurança pedonal.
Full text
Universidade do Minho Escola de Engenharia Francisco Emanuel Cunha Soares Risk factors affecting pedestrian behaviour: risk assessment in a virtual environment may 2022 UMinho | 2022 Francisco Emanuel Cunha Soares Risk factors affecting pedestrian behaviour: risk assessment in a virtual environment
Francisco Emanuel Cunha Soares Risk factors affecting pedestrian behaviour: risk assessment in a virtual environment Doctoral Thesis Doctoral Program in Civil Engineering Work conducted under supervision of Professor Doctor Elisabete Fraga de Freitas Doctor Emanuel Augusto Freitas de Sousa Universidade do Minho Escola de Engenharia may 2022
ii DIREITOS DE AUTOR E CONDIÇÕES DE UTILIZAÇÃO DO TRABALHO POR TERCEIROS Este é um trabalho académico que pode ser utilizado por terceiros desde que respeitadas as regras e boas práticas internacionalmente aceites, no que concerne aos direitos de autor e direitos conexos. Assim, o presente trabalho pode ser utilizado nos termos previstos na licença abaixo indicada. Caso o utilizador necessite de permissão para poder fazer um uso do trabalho em condições não previstas no licenciamento indicado, deverá contactar o autor, através do RepositóriUM da Universidade do Minho. Licença concedida aos utilizadores deste trabalho Atribuição CC BY https://creativecommons.org/licenses/by/4.0/
iii ACKNOWLEDGEMENTS This work was funded by Fundação para a Ciência e a Tecnologia (FCT) through the doctoral grant SFRH/BD/131638/2017, and it was incorporated in the research project “AnPeB - Analysis of pedestrians behaviour based on simulated urban environments and its incorporation in risk modelling” (PTDC/ECM-TRA/3568/2014), funded by the Promover a Produção Científica e Desenvolvimento Tecnológico e a Constituição de Redes Temáticas (3599-PPCDT) project and supported by the European Community Fund FEDER. In addition, the elaboration of this work would not have been possible without the contribution of a group of people and entities. In this way, I would like to express all my gratitude to all those who, directly or indirectly, contributed so that all this work had been developed. So, sincerely and particularly, I express my gratitude: To Professor Elisabete Freitas, for her guidance, support, availability, and readiness provided along the last years, and, above all, for all the knowledge conveyed to me. To Dr. Emanuel Sousa, for making the development of most of the tasks in this work more efficient with the link to the Centre for Computer Graphics (CCG), and for his support, orientation, promptness to offer his wise contributions and to transmit important skills to the execution of this project. To Professor Jorge Santos, for the important contributions provided in the first years of this project. His knowledge and experience were crucial to develop and close this entire project. To Professor Susana Faria, for her support in the data analysis and for the transmitted knowledge of statistics. To all the researchers, namely to Leidy Barón, Leandro Marcomini, Emanuel Silva, Carlos Silva, Sandra Mouta, Patrícia Barbosa, and Catarina Sousa, who have also contributed to the success of the AnPeB with their services and efforts. To João Lamas, Dário Machado, Frederico Pereira, Hélder Torres, and Carlos Palha, for the provided technical support and assistance.
iv To the Municipalities of Braga and Guimarães, for the support provided for the execution of the field observations. To JAPAutomotive – Renault Guimarães, for lending one of their electric vehicles to be used in recording the sounds implemented in the simulator. To all my friends, who, in a close or more distant way, accompanied me on this journey and beyond. To my family, specially to my parents, my brothers, and my girlfriend, for the motivation, the unconditional support, and the affection, which made this arduous journey smoother, and for everything they have always transmitted to me.
v STATEMENT OF INTEGRITY I hereby declare having conducted this academic work with integrity. I confirm that I have not used plagiarism or any form of undue use of information or falsification of results along the process leading to its elaboration. I further declare that I have fully acknowledged the Code of Ethical Conduct of the University of Minho.
vi FATORES DE RISCO QUE AFETAM O COMPORTAMENTO DE PEÕES: AVALIAÇÃO DO RISCO EM AMBIENTE VIRTUAL RESUMO A promoção de modos de transporte suaves, dos quais fazem parte andar a pé ou de bicicleta, ultimamente tem sido impulsionada devido às vantagens sociais e ambientais que estes possuem. Por outro lado, o aumento dos volumes de tráfego motorizado verificado nos últimos anos tem-se traduzido num aumento da exposição ao risco de acidente para todos os utilizadores da rede viária. No que se refere aos peões, esse crescimento leva a que ocorram mais interações entre estes utilizadores da rede viária e os veículos motorizados. O número de mortes de peões que ocorrem nas estradas está longe de ser nulo, sendo que uma parte considerável delas acontece em passagens para peões. Procurar prevenir a ocorrência de acidentes que possam ter graves consequências, promovendo, desta forma, maior conforto e segurança para todos os utilizadores da rede viária deve ser uma prioridade, tendo ciente que o paradigma da mobilidade está a mudar. Aproveitando as mais recentes tecnologias para a aquisição de dados sobre o comportamento de peões, como a análise automatizada de vídeo e o uso de um simulador onde num ambiente virtual se consegue estudar o comportamento dos peões quando executam o atravessamento da estrada sem que enfrentem um perigo real para a sua integridade física, o principal objetivo deste projeto de doutoramento foi a identificação e análise de fatores com influência no risco para peões quando atravessam a faixa de rodagem relacionados com as características das infraestruturas rodoviária e pedonal, as características do tráfego motorizado e pedonal e com o ruído emitido pelos veículos, dando particular relevância à tomada de decisão de atravessamento e à sua interação com os veículos. De uma forma geral, os resultados mostram a influência direta ou indireta dos diversos fatores abordados na tomada de decisão dos peões e na interação entre eles e os veículos que deles se aproximam aquando do atravessamento da faixa de rodagem. No entanto, os resultados que mais se evidenciam levam a concluir que a velocidade e a cinemática dos veículos em aproximação à passagem para peões são fatores com elevada importância para a segurança pedonal. Palavras-chave: Segurança rodoviária; Segurança pedonal; Decisão de atravessamento; Interação veículo-peão; Ambientes virtuais.
vii RISK FACTORS AFFECTING PEDESTRIAN BEHAVIOUR: RISK ASSESSMENT IN A VIRTUAL ENVIRONMENT ABSTRACT Soft modes of transportation, which include walking or cycling, have recently seen a boost in popularity and public promotion, due to their social and environmental advantages. This coincided with a growth of motorized traffic volumes in recent years, leading to more interactions between soft transportation users and motorized vehicles and greater exposure to the risk of accidents for all road users, with pedestrian being the most vulnerable ones. In fact, and despite a general trend of improvement in road safety, the number of pedestrian deaths occurring today is still substantial. A considerable part of them takes place in crosswalks. Identifying and analysing factors that may influence the behaviour of different road users is an important step in the process of designing changes to be made to the road infrastructure aiming to improve safety conditions. Trying to prevent accidents that could have serious consequences, and thus promoting greater comfort and safety for all road users, must be a priority, being aware that the mobility paradigm is changing with the emergence of different means of transport, namely electric vehicles. Taking advantage of the latest technologies to acquire data on pedestrian behaviour, such as automated video analysis and an augmented reality simulator where in a virtual environment it is possible to study the behaviour of pedestrians when crossing the road without putting their physical integrity in real danger, the main objective of this PhD project was the identification and analysis of factors influencing the risk for pedestrians when crossing the road, giving particular relevance to decision-making on crossing and interaction with vehicles. In general, the results show the direct or indirect influence of various factors and their interactions in the decision-making process of pedestrians when they cross a road. The most evident results lead to the conclusion that vehicle’s kinematics, especially its speed, are factors with high importance for pedestrian safety. Keywords: Road safety; Pedestrian safety; Crossing decision; Vehicle-pedestrian interaction; Virtual environments.
xiv Table 3.2 – Characteristics of vehicle movement in the different conditions presented in the experiment. .................................................................................................................. 54 Table 3.3 – Acoustic characteristics of the stimuli regarding the electric and gasoline combustion vehicles. ....................................................................................................................... 55 Table 4.1 – Mean, maximum, minimum, and standard deviation of vehicle speed and distance to pedestrian for each speed pattern. ................................................................................ 74 Table 4.2 – Characteristics of vehicle movement in the different conditions presented on the experiment. .................................................................................................................. 74 Table 4.3 – Descriptive statistics of the percentage of crossings, response time, and TTP for each trial and vehicle Constant speed pattern only. ................................................................ 79 Table 4.4 – Descriptive statistics of percentage of crossings, response time and TTP for each trial regarding all vehicle speed patterns. .............................................................................. 85 Table 4.5 – Summary of the main findings. ....................................................................................... 94 Table 5.1 – Participants’ demographic characteristics. ...................................................................... 99 Table 5.2 – Acoustic characteristics of the stimuli regarding the static approach (CPB sounds). ....... 100 Table 5.3 – Acoustic characteristics of the stimuli regarding the dynamic approach (CPX auralized sounds). ..................................................................................................................... 102 Table 5.4 – Descriptive statistics of percentage of crossings and crashes for each trial regarding the Constant speed pattern. ........................................................................................ 107 Table 5.5 – Descriptive statistics of crossing start time and TTP for each trial regarding the Constant speed pattern. ............................................................................................................ 108 Table 5.6 – Descriptive statistics of percentage of crossings and crashes for all the speed patterns. .................................................................................................................................. 115 Table 5.7 – Descriptive statistics of crossing start time and TTP for all the speed patterns. .............. 115
xv Table 6.1 – Participants’ demographic characteristics. .................................................................... 126 Table 6.2 – Main characteristics of each one of the six scenarios. ................................................... 127 Table 6.3 – Descriptive statistics of percentage of crossings, crossing start time and crashes registered by group of participants and respective street scenarios. ............................. 132 Table 6.4 – Statistical summary of the quantitative variables considered to model the TTP. ............. 137 Table 6.5 – Results of the 1st iteration of the TTP model. ................................................................. 139 Table 6.6 – Results of the final iteration of the TTP model. .............................................................. 140 Table 6.7 – Statistical summary of the quantitative variables considered to model the TTCmin. ........... 143 Table 6.8 – Results of the 1st iteration of the TTCmin model. ............................................................... 145 Table 6.9 – Results of the final iteration of the TTCmin model. ............................................................ 145
xvi LIST OF FIGURES Figure 1.1 – Research methodology and thesis overview. .................................................................. 14 Figure 2.1 – Example of the video camera’s placement (BMJ street, Braga). ..................................... 22 Figure 2.2 – Video frame from the six streets considered in Guimarães: (a) TP; (b) SG; (c) AGC; (d) G; (e) P; (f) L............................................................................................................ 23 Figure 2.3 – Video frame from the six streets considered in Braga: (a) 25A; (b) CL; (c) BMJ; (d) CJ; (e) GNM; (f) NSC. .......................................................................................................... 24 Figure 2.4 – Example of the video analysis made through Traffic Intelligence (BMJ street, Braga). ..... 25 Figure 2.5 – Frame from a video converted to MOG 2. ...................................................................... 26 Figure 2.6 – Example of the data exported by Traffic Intelligence: a) pedestrian and vehicle trajectories along time; b) vehicle speed along time; c) pedestrian speed along time. ..... 26 Figure 2.7 – Boxplot of TTP as a function of street. ........................................................................... 31 Figure 2.8 – Verification of the assumptions for the TTP model: (a) Q-Q plot; (b) Standardized residual versus fitted values. ......................................................................................... 36 Figure 2.9 – Verification of the assumptions for the CarSpeed_mean model: (a) Q-Q plot; (b) Standardized residual versus fitted values. ............................................................... 39 Figure 2.10 – Boxplot of TTCmin as a function of street. ....................................................................... 40 Figure 2.11 – Verification of the assumptions for the TTCmin model: (a) Q-Q plot; (b) Standardized residual versus fitted values. ......................................................................................... 44 Figure 3.1 – The two virtual scenarios: (a) 25A; and (b) TP. .............................................................. 52
xvii Figure 3.2 – Visual models of the vehicle used in the experiment: (a) Kia Ceed SW; (b) Renault Zoe ZE. ............................................................................................................................... 55 Figure 3.3 – Percentage of crossings and respective mean, standard error as a function of auditory condition, per vehicle speed pattern. ............................................................................. 58 Figure 3.4 – Percentage of cumulative crossings aggregated for all participants as a function of response time. .............................................................................................................. 60 Figure 3.5 – Distribution of the number of crossings with the mean standard error (horizontal lines) of the response time, by the participants, as a function of response time, per speed pattern and auditory condition. ...................................................................................... 61 Figure 3.6 – Mean TTP and respective mean standard error as a function of auditory condition, per vehicle speed pattern. ................................................................................................... 62 Figure 3.7 – Mean distance between the vehicles and participants at the moment of the participants’ responses and respective mean standard error as a function of auditory condition, per vehicle speed pattern. ............................................................................. 63 Figure 3.8 – Mean speed of the vehicle at the moment of the participants’ response and respective mean standard error as a function of auditory condition per vehicle speed pattern. ........ 64 Figure 4.1 – Comparison between the (a) virtual and (b) real scenarios regarding the 25A Street. ...... 72 Figure 4.2 – Comparison between the (a) virtual and (b) real scenarios regarding the TP Street. ........ 72 Figure 4.3 – Spatial layout of the room and participant position. ........................................................ 76 Figure 4.4 – Participants’ view during the performance of the experiment. ......................................... 77 Figure 4.5 – Participants’ responses along time of stimulus’ presentation, per auditory condition and initial speed (Constant speed pattern). .................................................................... 78 Figure 4.6 – Percentage of crossings and respective mean standard error as a function of auditory condition, per initial distance and initial speed (Constant speed pattern). ....................... 80
xviii Figure 4.7 – Mean response time and respective standard error as a function of auditory condition, per initial distance and initial speed (Constant speed pattern). ....................................... 81 Figure 4.8 – Mean values of TTP at the crossing instant as a function of the TTP at the start of the trial (Constant speed pattern). ....................................................................................... 82 Figure 4.9 – Participants’ responses and vehicle speed along time of stimulus’ presentation, per speed pattern, for initial speed of 20 km/h. ................................................................... 83 Figure 4.10 – Participants’ responses and vehicle speed along time of stimulus’ presentation, per speed pattern, for initial speed of 30 km/h. ................................................................... 84 Figure 4.11 – Percentage of crossings and respective mean standard error as a function of auditory condition, per initial speed and speed pattern. ............................................................... 86 Figure 4.12 – Mean response time and respective standard error as a function of auditory condition, per initial speed and speed pattern. ............................................................... 87 Figure 4.13 – Histogram of number of crossings by response time: (a) for the general data; and (b) by speed pattern and group of participants. .............................................................. 88 Figure 4.14 – Mean TTP and respective standard error as a function of auditory condition, per initial speed and speed pattern...................................................................................... 90 Figure 5.1 – Schematic representation of the auralization routine. ................................................... 101 Figure 5.2 – Visual model of the vehicle used in the experiment. ..................................................... 103 Figure 5.3 – Spatial layout of the room in the dynamic approach. .................................................... 104 Figure 5.4 – Participant performing the dynamic experiment. .......................................................... 104 Figure 5.5 – Participants’ responses along time of stimulus’ presentation, per initial distance, regarding the 20 km/h speed, for Constant speed pattern and for each experimental approach: (a) dynamic; and (b) static. ......................................................................... 106
xix Figure 5.6 – Participants’ responses as a function of the time of stimulus’ presentation, per initial distance, regarding the 30 km/h speed, for Constant speed pattern and for each experimental approach: (a) dynamic; and (b) static. ..................................................... 107 Figure 5.7 – Percentage of crossings and respective mean standard error as a function of experimental approach, per initial distance and initial speed, for Constant speed pattern. ...................................................................................................................... 110 Figure 5.8 – Crossing start time and respective mean standard error as a function of experimental approach, per initial distance and initial speed, for Constant speed pattern. ................. 111 Figure 5.9 – Mean values of TTP at the crossing instant as a function of the TTP at the start of the trial, by experimental approach, for Constant speed pattern. ........................................ 112 Figure 5.10 – Mean values of TTP at the crossing instant as a function of the TTP at the start of the trial, disregarding crashes, by experimental approach, for Constant speed pattern. .................................................................................................................................. 112 Figure 5.11 – Participants’ responses along time of stimulus’ presentation, regarding the initial speed of 20 km/h, per speed pattern and experimental approach: (a) dynamic; and(b) static. ......................................................................................................................... 113 Figure 5.12 – Participants’ responses along time of stimulus’ presentation, regarding the initial speed of 30 km/h, per speed pattern and experimental approach: (a) dynamic; and (b) static. .................................................................................................................... 114 Figure 5.13 – Percentage of crossings and respective mean standard error as a function of experimental approach, per initial speed and speed pattern. ........................................ 117 Figure 5.14 – Crossing start time and respective mean standard error as a function of experimental approach, per initial speed and speed pattern. ............................................................ 118 Figure 5.15 – TTP and respective mean standard error as a function of experimental approach, per initial speed and speed pattern.............................................................................. 120 Figure 5.16 – Comparison of the TTP obtained in real and virtual environment through the execution of each experimental approach, per speed pattern. ...................................... 121
xx Figure 6.1 – The six scenarios considered in this study: (a) 25A; (b) TP; (c) BMJ; (d) SG; (e) CL; (f) AGC. ...................................................................................................................... 128 Figure 6.2 – Participants’ responses along time of stimulus’ presentation, per group of participants and respective street scenarios: (a) Group I; (b) Group II; and (c) Group III. .................. 131 Figure 6.3 – Percentage of crossings and respective mean standard error as a function of street scenario, per group of participants. ............................................................................. 132 Figure 6.4 – Percentage of cashes and respective mean standard error as a function of street scenario, per group of participants. ............................................................................. 133 Figure 6.5 – Crossing start time and respective mean standard error as a function of street scenario, per group of participants. ............................................................................. 134 Figure 6.6 – Boxplot of TTP as a function of street scenario, per group of participants. .................... 135 Figure 6.7 – Boxplot of TTP without general dataset’s outliers, as a function of street scenario, per group of participants. .................................................................................................. 136 Figure 6.8 – Verification of the assumptions for the TTP model: (a) Q-Q plot; (b) Standardized residual versus fitted values. ....................................................................................... 140 Figure 6.9 – Boxplot of TTCmin as a function of street scenario, per group of participants. .................. 141 Figure 6.10 – Boxplot of TTCmin without general dataset’s outliers, as a function of street scenario, per group of participants. ............................................................................................ 142 Figure 6.11 – Verification of the assumptions for the TTCmin model: (a) Q-Q plot; (b) Standardized residual versus fitted values. ....................................................................................... 146
xxi GLOSSARY Time-to-passage (TTP): time remaining until an object ( e . g . vehicle) passes in front of an observer ( e . g . pedestrian), in seconds, if it continues with the speed and trajectory corresponding to the instant which the indicator is calculated (Hancock and Manser, 1998). The same as TTZ, TG, and T2 (Cavallo et al., 2019; Laureshyn et al., 2010; Lobjois and Cavallo, 2007; Várhelyi, 1998); Minimum time-to-collision (TTCmin): minimum time remaining until a collision occurs, in seconds, if, during the encounter, road users ( e . g . vehicle and pedestrian), continue with the speeds and trajectories that they had at the time for which the indicator was calculated ( e . g . the instant that a pedestrian starts to cross the road) (Archer, 2005; Hayward, 1972; Horst, 1990); Percentage of crossings: value resulting from the division between the number of crossings, i.e., the trials for which participants have clicked the computer mouse or started to cross the semi-virtual crosswalk before the vehicle has stopped or passed in front of them, and the total number of trials multiplied by one hundred. Its calculation was done per participant and for a given type of stimulus presented or a given movement condition of the approaching the vehicle in the experiences developed in a virtual environment; Response time: time, in seconds, from the start of the stimulus presentation to the moment the participant clicked the mouse in the static experimental approach; Crossing start time: time, in seconds, from the start of the stimulus presentation to the moment the participant started to cross the semi-virtual crosswalk in the dynamic experimental approach;
xxii Static approach: experimental approach where participants performed a road crossing task by clicking on a button and standing at a predefined position during all the experiment; Dynamic approach: experimental approach where participants performed a road crossing task walking along the semi-virtual crosswalk; CPB sounds: sounds regarding the movement of a vehicle recorded by an Head and Torso Simulator following the Controlled Pass-By (CPB) method. These sounds include all vehicle noise sources, the effect of propagation mechanisms, and noise from the surrounding environment (Freitas et al., 2012); CPX auralized sounds: sounds regarding the movement of a vehicle recorded by microphones type mounted on the back-right wheel of the vehicle following the Close Proximity (CPX) method. The signal captured by the CPX microphones is predominantly tyre-road noise. These sounds were then submitted to an auralization routine that outputs corresponding binaural CPB-like samples in order to allow a subject to hear a sound that appeared to come from an approaching vehicle.
1 1. INTRODUCTION 1.1. Background 1.1.1 Pedestrian fatalities in Portuguese and European roads Road safety is today a major concern of many regulatory and governing entities through the world. This is motivated by the severe social and economic impacts resulting from traffic accidents and most importantly from road deaths and injuries they cause. The number of road fatalities has decreased in the European Union (EU) during the last 20 years as a result of a myriad of measures such as infrastructure improvements, wiser regulations and a constant demand for better safety features on the vehicles. However, according to the EU (European Commission, 2018b), the downward trend in the percentage of vulnerable road users’ fatalities, particularly pedestrians, is not evolving as other indicators. In the decade between 2007 and 2016, pedestrian fatalities decreased by 36 %, while the total number of road fatalities decreased by almost 41 % (European Commission, 2018b). The proportion of pedestrian traffic fatalities was still 21% in 2018 (CARE, 2020b; European Commission, 2018a), compared to the total of deaths occurred in the European roads. In Portugal, the downward trend was not as pronounced as in the group of the EU countries. The pedestrian fatalities were reduced by 10 %, while road fatalities decreased by 35 %. Furthermore, the percentage of pedestrian fatalities was above the European average. In 2016, 22 % of all deaths in the Portuguese roads were pedestrians (ANSR, 2007, 2016; European Commission, 2018b). That percentage had a small decrease in 2019 when 21 % of all road deaths were pedestrians (ANSR, 2019). Coupled with those numbers, the ratio of 13.01 pedestrian fatalities per million population registered in 2019, higher than the European average, shows that Portugal is a country where there are still problems related to pedestrian safety (see Table 1.1). While the country fares substantially better than other European countries, such as Romania and Latvia, were 35.33 and 25.85 pedestrian fatalities per million people were reported, it is still far from the best placed ones, such as Netherlands and Sweden, where only 2.88 and 3.32 pedestrian fatalities per million population were registered (ANSR, 2019; CARE, 2020b).
INTRODUCTION 8 al. (2021) identified a new group of methods in different reviews of data collection methods used in studies about pedestrian behaviour: the controlled experiments, which may be split into naturalistic experimentation and simulation. Regarding pedestrians’ field observations when crossing the road, they can be used to get rich information about road users’ movement and behaviour. However, since this method does not involve all variables’ control, many observations are necessary, which can make data collection and analysis very timeconsuming (Feng et al. , 2021). The most common way of gathering this data is through video recordings of crossing situations (Lassarre et al. , 2012). Video automated or semi-automated processing and analysis tools have lately started to turn field observations into a more efficient method ( e . g . (Ismail et al. , 2009; Jackson et al. , 2013; Johnsson et al. , 2018b; Saunier et al. , 2010)). As referred by Olszewski et al. (2016), these tools are habitually used due to their versatility, low cost, and the content of the data they can export. However, video recordings are limited to the camera’s field of vision. Usually, the road users are detected in an area of 30 – 40 m wide. High levels of luminosity, fog, rain, snow, and nighttime, may jeopardize or even prevent the data collection (Olszewski et al. , 2016). Other alternatives to video analysis are following pedestrians’ trajectory through GPS instruments or Bluetooth/Wi-Fi sensors. These tools present some limitations compared to video analysis, namely problems with precise location, unavailability of information regarding traffic conditions, and they can influence the users’ behaviour, since they imply pedestrians’ instrumentation (Feng et al. , 2021; Papadimitriou et al. , 2016a). As said by Deb et al. (2018a), surveys can be written documents, online questionnaires, face-to-face interviews, or telephone interviews. This type of method allows for great controllability of all the variables inserted in a study since researchers’ questions in a survey are previously design. However, the answers given by participants may not portrait their actions in real situations. This method is useful to complement the data gathered in field observations or controlled experimentation because it provides the opportunity to acquire information about pedestrians’ personal and psychological characteristics (Feng et al. , 2021). Controlled experiments comprise the most used semi-controlled naturalistic experiments and virtual reality experiments (Deb et al. , 2018a; Feng et al. , 2021; Kircher et al. , 2017). Beyond many advantages, both have the disadvantage that they may cause participants’ behaviour to be unrealistic since they aware that they are being observed and analysed, such as in the usage of surveys (Feng et al. , 2021).
CHAPTER 1 9 According to Kircher et al. (2017), semi-controlled studies follow a hybrid approach between controlled experiments and observational methods. In semi-controlled naturalistic experiments, the researcher can previously define groups of participants, routes, and tasks. However, there is a set of uncontrolled variables present, for instance, vehicle speeds, traffic volumes, pedestrian and motorized traffic densities, etc. This method is usually applied to analyse factors such as gait parameters and pedestrian spatial organization in real environments (Cao et al. , 2018; Fu et al. , 2019; Wei et al. , 2015). The relatively high controllability of variables makes this experimental method the most effective in providing complete information to analyse particular factors (Feng et al. , 2021; Kircher et al. , 2017). However, the lack of precise control of all the variables involved in the study can turn the analysis of those factors complicated (Kircher et al. , 2017). In addition, this method requires setting up data collection devices, which is labour intensive (Feng et al. , 2021). Virtual reality experiments have been used in situations where real-world environments are difficult to control or dangerous (Deb et al. , 2017). Schwebel et al. (2008), citing Reid (2002), define virtual reality “as a computeror video-generated environment that gives the user a sense of being in a displayed virtual world through realistic images, high-quality sound, the feeling of immersion, and the ability to interact with the virtual world”. The use of this tool in controlled experiments allows the simulation of the more diverse situations, where all the variables can be easily controlled since the virtual scenes can be quickly built and modified (Deb et al. , 2017; Feng et al. , 2021). Virtual reality experiments allow the recreation in controlled experimental settings of risky situations, such road crossings, but without exposing the participants to a real risk (Deb et al. , 2017; Schwebel et al. , 2008). Compared to the other data collection methods, this allows for a more accurate data collection due to the mentioned controllability. Furthermore, that data can be easily processed and quickly analysed (Feng et al. , 2021). However, simulators are expensive in set-up and maintenance. They require custom-built solutions (Deb et al. , 2017), which need sufficient space to consider participants’ movement during the experiments. 1.1.2.4 Risk factors The risk factors associated with pedestrian behaviour and safety identified in the literature are vast. But, regarding the ones more often addressed, they are essentially distributed into three distinct groups: those
INTRODUCTION 10 concerning the characteristics of pedestrians themselves, those regarding the characteristics of the road and built environment, and the factors related to traffic characteristics. Within the group of pedestrian characteristics, pedestrians’ age and sex have been the most addressed factors. Some authors refer that young pedestrians are more likely to make unsafe crossing decisions than older pedestrians because of the lack of experience, unpredictable behaviours, distraction, and poor risk perception (Bernhoft and Carstensen, 2008; Ezzati Amini et al. , 2019; Holland and Hill, 2007; Johansson et al. , 2004; Moyano Dıaz, 2002; Rosenbloom et al. , 2008). Regarding pedestrians’ sex, the conclusions are not consensual. However, when differences between female and male risk-taking behaviours are found, females are generally more conservative than males, acting more safely (Hamed, 2001; Holland and Hill, 2007; Moyano Dıaz, 2002; Papadimitriou et al. , 2016b). A few studies have also reported effects of pedestrians’ cultural, socioeconomic, and educational profile. However, these factors are difficult to analyse properly as such analysis requires considerable data collection of the same indicators at different points of the globe or a comparison of studies with similar methods and approaches to obtain reliable results (Sueur et al. , 2013). Concerning the characteristics of the road and built environment, road width, the number of lanes, the function of the surrounding buildings, the width and quality of the sidewalks, the marked parking places, and various pedestrian engineering and crossing treatments are factors which some authors argue that may directly or indirectly influence the pedestrian safety and their crossing interaction with vehicles (Ewing and Dumbaugh, 2009; Granié et al. , 2014; Lin et al. , 2015; Sucha et al. , 2017; Turner et al. , 2006; Zegeer et al. , 2006). Pedestrians may feel safer when the road width is shorter and the number of lanes is lower (Sucha et al. , 2017; Turner et al. , 2006; Zegeer et al. , 2006). In contrast, the lack of shops and the small number of houses, tight sidewalks or lateral space dedicated to pedestrians, the inexistence of marked parking places in the area involving the crosswalk are all factors that lead pedestrians to feel uncomfortable and to consider unsafe to cross the road. With all of these features which usually characterize an unattractive zone to walk, pedestrians infer a low density of pedestrian traffic, which they relate to the existence of better conditions to the practice of higher speeds by drivers (Granié et al. , 2014). Pedestrian and motorized traffic characteristics can influence pedestrian safety, impacting from pedestrian crossing decision-making to the occurrence of accidents. According to Ezzati Amini et al. (2019) and Sucha et al. (2017), pedestrians consider traffic density a factor when making the crossing
CHAPTER 1 11 decision. High motorized traffic volumes are considered riskier to pedestrians since the likelihood of accident and their injury severity, increases with the number of vehicles passing by a determined placed (LaScala et al. , 2000; Leden, 2002; Papadimitriou et al. , 2012). Additionally, drivers’ tendency to yield at crosswalks also decreases with high motorized traffic volumes (Sucha et al. , 2017). Pedestrian traffic volumes have a contrary effect on their safety. According to Leden (2002), the risk of an accident involving pedestrians decreases with increasing pedestrian flows. The author argues that one explanation could be the higher driver alertness about the presence of pedestrians. However, if not accompanied by appropriate traffic and safety conditions, the higher pedestrian traffic volumes can lead to more pedestrian accidents (Leden, 2002). On the other hand, for some authors, the vehicle approaching speed is one of the most important factors, if not the most, used by pedestrians to explain their crossing decision (Granié et al. , 2014; Sucha et al. , 2017). According to Várhelyi (1998), the vehicle speed is the unique single factor relevant to pedestrian safety and feeling of safety. It can affect their crossing decision and behaviour, since a fast-approaching vehicle can pressure pedestrians and might force them to cross the road unsafely. Furthermore, although pedestrians are aware that due consideration should be given to vehicle speed when making crossing decisions, they may not perceive the vehicle speed and misjudge the gap time available to cross the road (Liu and Tung, 2014). Other factor which has been attracting the efforts of some researchers is the vehicle noise. People are used to vehicles making noise because until very recently there were only vehicles with combustion engines in worldwide roads. With the appearance of electric and hybrid vehicles, noise levels emitted are reduced (Verheijen and Jabben, 2010; Wogalter et al. , 2001). What, on the one hand, can be considered a positive aspect given the negative impact that traffic noise has directly or indirectly on human health, on the other hand, it can be a serious problem from the point of view of road safety. Vehicular noise often acts as a cue for vulnerable road users, helping them to detect and locate the approaching vehicles and improving their perception of speed and distance. Thus, lower noise emissions can put pedestrian safety at risk, particularly blind or vision impaired ones (Barton et al. , 2013; Barton et al. , 2012; Emerson et al. , 2013; Emerson et al. , 2011; Verheijen and Jabben, 2010; Wiener et al. , 2006).
INTRODUCTION 12 Appendix I presents a table summarizing the data collection methods, the safety and behavioural indicators, and the factors addressed in all the studies about pedestrians’ crossing safety cited throughout this document. 1.2. Objectives As previously exposed, the numbers of pedestrian fatalities that occur on the roads, particularly in Portugal and throughout the European Union, are worrying. Therefore, questions about pedestrian safety, the interaction of pedestrians with motorized traffic, and their operational effects need to be explored. In this sense, understanding the pedestrian behaviour in urban environments, especially when crossing the road, is an essential issue to achieve the desired purposes of "Vision Zero". Taking advantage of the latest technologies for the data acquisition on pedestrian behaviour, such as automated video analysis and the use of simulators and the previously mentioned indicators, the general objective of this doctoral project was the identification and analysis of risk factors for pedestrians when crossing the road, with a particular focus on their crossing decision-making and interaction with vehicles approaching the crosswalk. This document is composed of several chapters that explore gaps in the existing literature. Those gaps are identified in the introduction section of each one of them. Nevertheless, according to the data collection methodology about the pedestrians’ behaviour at road crossings, the work can be divided into two distinct parts. One part based on field observation, which was performed through video recordings; and another based on controlled experiments carried out in a virtual environment. In addition to the transversal objective of this doctoral project, the field observation aimed to collect data on the vehicles’ movement approaching the crosswalk to achieve the following primary objectives of this work: the construction of the simulator to be used in controlled experiences; and to identify vehiclepedestrian interactions risk factors related to the pedestrians’ demographic, the road and pedestrian infrastructure, and the pedestrian and motorized traffic characteristics. In this way, the weights of these factors can be evaluated after a careful selection, using a more in-depth analysis performed in a virtual environment.
CHAPTER 1 13 The goals of the controlled experiments were to evaluate how the noise emitted by the approaching vehicles and their approaching movement pattern can affect the pedestrians’ crossing decision-making, and to complement the identification of factors associated with the pedestrians’ demographics, the road, and the pedestrian infrastructure characteristics, not neglecting the speed of approach of vehicles, influencing the pedestrians crossing decision-making, as well as their interaction with vehicles, carried out in the real environment part of this study. 1.3. Outline The conducted scientific research work described on this thesis is organized into seven chapters. The overall research methodology is presented in Figure 1.1 and it incorporates three main tasks: (i) literature review; (ii) analysis of pedestrian behaviour in a real environment; and (iii) analysis of pedestrian behaviour in a virtual environment. The present chapter (Chapter 1) contains a brief background and review of pedestrian fatalities in EU and Portugal, of the safety and behavioural indicators, the data collection methods usually, and the risk factors more often addressed and used in pedestrian road crossing safety studies. This chapter aims to support the introduction and the methodology sections of Chapter 2, 3, 4, 5 and 6. Then, it specifies the work’s main objectives and ends with a summary of the content of this thesis. Chapter 2 presents a study carried out in a real environment where data from vehicle-pedestrian interaction were collected through video recordings in twelve different crosswalks. The pedestrian crossing decision-making and the severity of the encounters between them and vehicles were analysed in terms of TTP and TTCmin, respectively. Given the wide range of risk factors identified in the literature, the aim of the analysis carried out in this chapter was to identify risk factors to vehicle-pedestrian interactions related to the pedestrians’ demographics, the road and pedestrian infrastructures, motorized and pedestrian traffic characteristics, during road crossings at unsignalized crosswalks, in order to select them for a more in-depth evaluation in the following chapters. A special focus was given to vehicles’ approaching speed. The data collected about vehicles’ trajectories and speeds were used to configure the virtual scenarios used in the simulator (Chapter 3, 4, 5, and 6).
INTRODUCTION 14 Figure 1.1 – Research methodology and thesis overview. Virtual Environment Chapter 6 - Performance of an experiment, following the approach implemented in the study presented on the previous chapter, but using two new pairs of virtual scenarios, considering two sets of 15 participants; - Complementing the conclusions of Chapter 2 on the effect of the pedestrian demographics, the built environment, and the traffic characteristics on pedestrian crossing decision-making and vehicle-pedestrian interaction. Chapter 3 - Exploring the analysis of pedestrian behaviour on the developed simulator with the execution of an experiment to a set of 10 participants, using two virtual scenarios; - First analysis of the effect of noise (electric engine, gasoline combustion engine, and no sound) and vehicles’ approaching movement on pedestrian crossing decision-making. Chapter 4 - Extension of the study presented on the previous chapter with the execution of a different experiment to a set of 30 participants, using two virtual scenarios, in order to get more robust results; - Analysis of the effect of the auditory cues (gasoline combustion engine and no sound) on pedestrian crossing decision-making; - Analysis and clarification of the effect of vehicle’s approaching kinematics on pedestrian crossing decision-making. Chapter 5 - Execution of an experiment following a different approach, where 30 participants performed a road crossing walking along a semi-virtual crosswalk in two virtual scenarios; - Comparison between both experimental approaches used in the simulator described in Chapter 4 and Chapter 5. Chapter 7 - Main conclusions. - Recommendations for future works. Real Environment Chapter 2 - Assessment of the effect of the pedestrian demographics, the built environment, and the traffic characteristics on pedestrian crossing decision-making and vehicle-pedestrian interaction through the analysis of 2 hours video recordings performed at 12 study sections; - Data acquisition on vehicles’ trajectories and speeds for implementation on the simulator. Literature Review Chapter 1 - Background on the pertinence of the thesis topic and on the data collection methods and the behavioural/safety indicators addressed in studies on pedestrian safety. - Main objectives and outline.
CHAPTER 1 15 Chapter 3 comprises a first approach to analyse pedestrian behaviour using the developed simulator. Given the relevance that electric vehicles and traffic noise have received in recent times and the existence of a literature gap regarding the simulation of the movement of vehicles approaching the crosswalk in studies using a virtual environment to assess pedestrians’ behaviour and safety, this chapter aimed to analyse the influence of vehicular noise, as well as vehicle’s approaching speeds and trajectories, on pedestrian crossing decision-making. The importance of the vehicle’s approaching movement and the auditory cues on pedestrian crossing decision-making was analysed in terms of percentage of crossings, response time, and TTP. Three auditory conditions, three vehicle’s movement conditions, and two scenarios were depicted in the ninety stimuli presented to a reduced sample of ten participants which performed an experimental task of a road crossing situation standing at a predefined position and clicking in a computer mouse when intended to cross the virtual road. This experimental approach was called “Static approach”. Chapter 4 consists of an extension of the study presented in Chapter 3, intending to confirm its results through a more robust data sample. Considering the gaps in the literature and the reduced relevance given to factors such as the auditory cues and the vehicle’s approaching movement as a starting point, the aim of the work presented in the Chapter 4 was to clarify the role of the vehicle kinematics on pedestrians’ crossing decision-making, mediated by the resulting visual and auditory cues. A more thorough analysis of the role of speed and distance, as well as the different speed profiles of the approaching vehicle was performed. Such as in Chapter 3, the effect of the approaching movement and the noise emitted by the vehicle on pedestrian crossing decision-making was also analysed in terms of percentage of crossings, response time, and TTP. Two auditory conditions, ten vehicle’s movement conditions, and two scenarios were depicted in the two-hundred stimuli presented to thirty participants. Chapter 5 describes a methodological study where two experimental approaches are compared: (i) the approach used in Chapter 3 and Chapter 4, where participants performed a road crossing task by clicking on a button and standing at a predefined position during all the experiment (Static approach); (ii) an approach where participants performed the same task walking along the virtual crosswalk, called “Dynamic approach”. The aim of this analysis was to evaluate the impact of the implementation of a more realistic approach for studying pedestrian crossing behaviour and to assess the advantages or disadvantages of the use of each approach, since both are more frequently used in pedestrian safety studies. The data collected in the study presented in Chapter 4 from the thirty participants was compared
INTRODUCTION 16 to that collected from another thirty participants who performed the experiment following a different approach in terms of percentage of crossings, response time, and TTP. Subsequently, the TTP obtained in the virtual environments was compared to those obtained in a real environment. Chapter 6 is based on an extension of the virtual environment of the study presented in Chapter 2. The experimental protocol related to the dynamic experience introduced and presented in the study described on Chapter 5 was considered to assess the risk factors to vehicle-pedestrian interactions related to the pedestrians’ demographics, the road and pedestrian infrastructures, motorized and pedestrian traffic characteristics, during road crossings in order to complement the findings of the analysis of pedestrian behaviour in the real environments (Chapter 2). The pedestrian crossing decision and the severity of the encounters between them and the vehicle were analysed in terms of percentage of crossings, crossing start time, percentage of crashes, TTP, and TTCmin. Fifty stimuli were presented in six different scenarios to a total of forty-five participants. Although the conclusions are included in each chapter, a summary of the work carried out and its main conclusion are presented in Chapter 7. It is also given a set of recommendations for future works.
17 2. ANALYSIS OF PEDESTRIAN BEHAVIOUR IN REAL ENVIRONMENT 2.1. Introduction During the last years, governing bodies in several countries have been trying to improve road safety, by raising awareness of road users and thus trying to improve behaviours or by making changes to road infrastructure. Greater attention is provided to users of soft modes of transport, which include cyclists and pedestrians, since they are the most vulnerable to road accidents. Still, accidents continue to be frequent and more needs to be done. Identifying and analysing factors that may influence road users’ behaviour is an essential step in designing road infrastructure changes and public policies to improve safety conditions. Regarding pedestrian safety, exploring the interaction of these road users with the motorized traffic is key to understand what may affect their safety and put their integrity and life at risk. Several studies have been carried out addressing pedestrian behaviour, especially during road crossing situations, because most of the accidents involving pedestrians happen in those situations. Crosswalks are places where a bigger number of conflicts between pedestrians and road traffic occurs and, thus, where these vulnerable road users are more exposed to the risk of accident (Lassarre et al. , 2007). Following that which is described in Chapter 1, the risk factors associated with pedestrian behaviour and safety are vast. They are distributed into three distinct groups: those concerning the pedestrians’ characteristics, those regarding the characteristics of the road and built environment, and the factors related to traffic characteristics. Within the group of pedestrian characteristics, pedestrians’ age and sex have been the most addressed factors (Bernhoft and Carstensen, 2008; Ezzati Amini et al. , 2019; Hamed, 2001; Holland and Hill, 2007; Johansson et al. , 2004; Moyano Dıaz, 2002; Papadimitriou et al. , 2016b; Rosenbloom et al. , 2008). Pedestrians’ cultural, socioeconomic, and educational profile are other factors sometimes considered in pedestrian safety studies (Sueur et al. , 2013).
ANALYSIS OF PEDESTRIAN BEHAVIOUR IN REAL ENVIRONMENT 24 (a) (b) (c) (d) (e) (f) Figure 2.3 – Video frame from the six streets considered in Braga: (a) 25A; (b) CL; (c) BMJ; (d) CJ; (e) GNM; (f) NSC.
CHAPTER 2 25 2.2.3 Video analysis Video analysis was carried out through the Traffic Intelligence program (Ismail et al. , 2009; Jackson et al. , 2013; Saunier et al. , 2010) (Figure 2.4). In essence, for each recorded video, the program identified and mapped the motion of each pixel, frame by frame, grouped them by the similarity of characteristics, and classified each group as an object (pedestrian or vehicle), thus obtaining the cars and pedestrians’ trajectories and speeds. Figure 2.4 – Example of the video analysis made through Traffic Intelligence (BMJ street, Braga). Other tasks were inherent to the process of video analysis, such as camera calibration, which consists of extracting the matrix and calculating its distortion coefficients, and the points homography, which corresponded to associating the video coordinates with the real-world coordinates. Before the tasks previously mentioned, the videos needed to be converted into MOG 2. MOG 2 creation process consists of a background removal that calculates the foreground mask by performing a subtraction between the current frame and a background model, containing the static part of the scene or, more generally, everything that can be considered as background given the characteristics of the observed scene. An example is presented in Figure 2.5. Since the video analysis process was a very time-consuming task which included performing subtasks for each video, such as homography, MOG 2 conversion, software calibration, processing the analysis itself, and refining and cleaning the exported data, and due to the fact of some sections have low pedestrian traffic volumes, the data collection from each street was limited to fifty crossing movements to avoid a large decompensation of the statistical weight of the variables related to the characteristics of the study
ANALYSIS OF PEDESTRIAN BEHAVIOUR IN REAL ENVIRONMENT 26 sections. The total number of observations considered in this study was 459 pedestrian-vehicle encounters. Figure 2.5 – Frame from a video converted to MOG 2. An example of the data collected with the video analysis made through Traffic Intelligence is presented in Figure 2.6. (a) (b) (c) Figure 2.6 – Example of the data exported by Traffic Intelligence: a) pedestrian and vehicle trajectories along time; b) vehicle speed along time; c) pedestrian speed along time. 0 10 20 30 010 20 30 40 Y (m) X (m) 0 10 20 30 0 3 6 9 12 15 Vehicle speed (km/h) Time (s) 0 10 20 30 0 3 6 9 12 15 Pedestrian speed (km/h) Time (s)
CHAPTER 2 27 2.2.4 Linear mixed-effects model The analysis of pedestrian crossing behaviour was carried out using two safety indicators as dependent variables, TTP and TTCmin. The first one allows to evaluate the severity of the encounter, i.e. , the possibility of an accident to occur, and the last one allows the assessment of the risk taken by pedestrians when they started the crossing task. PET was not used, since it requires consideration of the dimensions of the vehicles and pedestrians involved, which are not provided by the automated video analysis program. The impact of several variables related to the road and pedestrian infrastructure characteristics, the pedestrian demographic characteristics, and the pedestrian and motorized traffic characteristics on pedestrians’ risk in crossing decision-making and during the effective crossing were then assessed considering both variables. TTCmin is the minimum time remaining, during the encounter, before a collision occurs if road users, vehicle and pedestrian, continue with the speeds and trajectories that they had at the time for which the indicator was calculated (Archer, 2005; Hayward, 1972; Horst, 1990) (see Expression 2.1). For its determination, pedestrian and the vehicle’s movement were considered during the time interval from the beginning of the crossing until the moment when one of them passed the intersection point between the trajectory of the vehicle and the pedestrian, i.e. , the point of conflict. TTCmin = min [0; t] (max (Dvehicle-conflict point, i / Vvehicle, i; Dpedestrian-conflict point, i / Vpedestrian, i), i ϵ [0, …, t] (2.1) Where: - t is the encounter duration; - Dvehicle-conflict point is the distance, in m, from the centre of the vehicle’s license plate, assumed as the possible point of impact, and the point of conflict; - Vvehicle is the vehicle speed, in m/s; - Dpedestrian-conflict point is the distance, in m, from the pedestrian to the point of conflict; - Vpedestrian is the pedestrian speed, in m/s.
ANALYSIS OF PEDESTRIAN BEHAVIOUR IN REAL ENVIRONMENT 28 TTP is a psychophysical indicator that consists of the time remaining until an object (vehicle) passes in front of an observer (pedestrian) if it continues with the speed and trajectory corresponding to the instant which the indicator is calculated (Hancock and Manser, 1998) (see Expression 2.2). In this case, this instant corresponded to the moment when the pedestrian crossing started. This indicator corresponds to TTZ, TG, and T2 used by some authors (Cavallo et al. , 2019; Laureshyn et al. , 2010; Lobjois and Cavallo, 2007; Várhelyi, 1998). TTP = Dvehicle-conflict point / Vvehicle (2.2) Where: - Dvehicle-conflict point is the distance, in m, from the centre of the vehicle’s license plate, assumed as the possible point of impact, to the point of conflict at the moment when pedestrian started to cross; - Vvehicle is the vehicle speed, in m/s, at the moment when the pedestrian started to cross. In some encounters, it was impossible to get the value referring to one of the two indicators due to the incapacity to detect all the vehicles’ trajectories when pedestrians started the crossing. The final dataset thus contained, 285 observations for the TTP analysis and 459 for the TTCmin. In addition to the characteristics of the pedestrian and road infrastructure presented in Table 2.2 and Table 2.3, the pedestrian and motorized traffic volumes, shown in Table 2.4 and Table 2.5, and the demographic characteristics of pedestrians, age, and sex, both databases also include the average vehicle speed gathered during each encounter. As the number of observations among the groups compounding the variables was excessively unbalanced, the road directions (one or two directions), road pavement (asphalt concrete or cobble stones), and the number of lanes (two or four) were disregarded. A special relevance was given to vehicle approaching speed due to the TTP model results and the conclusions of some studies referred in the introductory section. Linear mixed-effects models (LMM) were used to assess the influence of all the variables previously mentioned on TTP and TTCmin. LMM extend linear models by incorporating random effects, which can be
CHAPTER 2 29 regarded as additional error terms, to explain the correlation between observations within the same group. The LMM expresses for the ith subject as (Pinheiro and Bates, 2006): Yi = Xi β + Zi bi + εi, i =1, …, n (2.3) Where: - Yi ϵ (Yi1 … YiTi)T represents a vector (of size Ti × 1) of continuous responses for the ith subject; - Xi is the known fixed-effects covariates matrix (of size Ti × p); - β is a vector (of size p × 1) of unknown regression coefficients (or fixed-effects parameters); - Zi is the known random-effects covariates matrix (of size Ti × q); - bi is a vector (of size q × 1) of random-effects; - εi represents an error vector (of size Ti × 1) of n residuals associated with an observed response for the ith subject. Moreover: bi ~ Nq (0, D) (2.4) εi ~ NTi (0, Ri) (2.5) Where: - D is the q × q covariance matrix for the random effects, and Ri is the Ti × Ti covariance matrix of the errors in group i; - bi e εi are independent for the same ith subjects and of each other.
ANALYSIS OF PEDESTRIAN BEHAVIOUR IN REAL ENVIRONMENT 30 Following the methodology described on Pinheiro and Bates (2006), the significance of fixed effects’ terms in the model was assessed by conditional F-tests using a sequential sum of squares. The Restricted Maximum Likelihood (REML) was used to estimate the parameters of the model. Akaike’s Information Criterion (AIC) and Bayesian Information Criterion (BIC) were used to choose the most appropriate structure for the covariance matrix of the errors, and the independent structure was selected (this structure assumes a homogeneous residual variance for all observations). LMM is an appropriate technique for analysing nested structured data, such as the data presented in this study. The encounters are nested within streets, and repeated measures were collected on twelve different streets. The applicability of this technique was previously assessed before the construction of the models. When LMM was not applied, a simple linear regression model was considered. This work’s modelling approach followed the backward deletion method, which consisted of iteratively removing the statistically non-significant variable with the highest p‑value. The final model for each response variable, namely TTP and TTCmin, presents the explanatory variables that were statistically significant to a level of 5 %. Before the first iteration of the model’s construction task, a Pearson correlation analysis was performed to examine whether some explanatory variable was highly correlated with the response variable. All statistical analyses were performed using the R statistical software (R Core Team, 2020). 2.3. Results 2.3.1 Time-to-passage An one-way ANOVA revealed statistically significant differences in TTP between the streets, F(11, 273) = 2.08, p = 0.02. Analysing Figure 2.7, it is possible to note that the registered TTP not only had considerable variations between streets, with the higher and lower mean values having been registered to the TP (m = 4.59 s; sd = 1.93 s) and GNM (m = 2.90 s; sd = 1.06 s) streets, respectively, but also within the street itself, particularly in NSC (m = 3.58 s; sd = 2.01 s), TP, and CL (m = 3.91 s; sd = 1.92 s). The statistical summary of the quantitative variables used in the modelling analysis of TTP presented in Table 2.7 shows that, in general, the TTP ranged from 1.02 to 8.92 s (m = 3.69 s; sd = 1.73 s). It is also possible to notice that the difference between the maximum value of TTP (Max) and the value of the third quartile (Q3) is bigger than the difference between the value of the first quartile (Q1) and its minimum value
CHAPTER 2 31 (Min), indicating a positively skewed distribution of the TTP. This situation was expected because TTP is a temporal indicator, limited to positive values, which forces a boundary in the zero value. This feature could cause the violation of one of the assumptions that must be met for applying the LMM, which is the normality of the distribution of residuals. A new variable was created based on its natural logarithm (logTTP) to lead with this instead of considering the TTP as the model’s dependent variable. The mean speed of the vehicles throughout the encounters ranged from 2.13 to 50.69 km/h (m = 17.88 km/h; sd = 9.22 km/h). In average terms, lower vehicle approaching speeds were registered in CJ (m = 7.96 km/h; sd = 2.76 km/h) and TP (m = 10.70 km/h; sd = 5.03 km/h) streets. In turn, SG (m = 27.20 km/h; sd = 11.60 km/h) and CL (m = 23.40 km/h; sd = 9.48 km/h) were the streets with the highest recorded values. Figure 2.7 – Boxplot of TTP as a function of street. Table 2.6 presents the statistical summary of the quantitative variables used in the modelling analysis of TTP. Regarding the pedestrians’ age, most of the sample comprises data gathered from road crossings done by adult pedestrians, i.e. , those belonging to the [20 – 40[ and [40 – 60[ age groups. Despite the sample’s balance, more crossings were observed of female (55.09 %) than male pedestrians. 70.88 % of all the considered pedestrian-vehicle encounters took place on a crosswalk located on a collector road. Regarding the remaining variables, both those related to the road and pedestrian infrastructure characteristics and traffic volumes, the minimum and maximum values presented in Table 2.7 are in line with those previously presented in Table 2.2, Table 2.3, Table 2.4, and Table 2.5, contrary to the other
ANALYSIS OF PEDESTRIAN BEHAVIOUR IN REAL ENVIRONMENT 32 values presented (Q1, Q3, Median, Mean, SD and Coefficient of variation) since they depended on the number of pedestrian-vehicle encounters recorded in each street. Table 2.6 – Frequencies of the qualitative variables considered to model the TTP. Variable Abbreviation Group Absolute Frequency Relative Frequency (%) Pedestrian’s age PartAge < 20 30 10.53 [20 – 40[ 135 47.37 [40 – 60[ 74 25.96 > 60 46 16.14 Pedestrian’s sex PartGen Female 157 55.09 Male 128 44.91 Road classification Road_class Collector 202 70.88 Local 83 29.12 The Pearson correlation analysis revealed that the mean speed of the vehicles along the encounters was the variable with the highest correlation with TTP (ρ = - 0.22), even though this was a low value (|ρ| < 0.30) (Hinkle et al. , 2003). Following, as a first step, the null model, i.e. , the model with no covariates, was fitted (Expression 2.6). The null model is useful for deciding whether a random-effects model might be appropriate for the data. Since σP2 = 0.0085 and σε2 = 0.2178, only 3.75 % (0.0085 / (0.2178 + 0.0085)) of the data variation is explained by allowing the intercept to vary across the streets, indicating that unobserved heterogeneity of logTTP among the streets may not be captured by using a random-intercept model. Thus, a simple linear regression was fitted to identify the variables with a significant effect on logTTP. The first iteration of modelling task considered all the variables previously described (see Expression 2.7). logTTPs,i = β0,i + b0,i + εs,i, (2.6) b0,i ~ N (0, σP2), εs,i~ N (0, σε2), s = 1st, …, 12th street, and i = 1st, …, 285th observation
CHAPTER 2 33 Table 2.7 – Statistical summary of the quantitative variables considered to model the TTP. Coef. of Variation 0.47 0.40 0.52 0.22 0.11 0.89 0.54 0.21 0.67 0.43 0.91 SD 1.73 0.48 9.22 1.88 0.36 3.42 1.81 0.84 42.17 300.41 170.16 Max 8.92 2.19 50.69 13.80 3.87 10.12 6.57 5.35 195.20 1286 631 Q3 4.54 1.51 23.46 9.96 3.57 5.17 4.17 4.48 104.30 784 276 Mean 3.69 1.20 17.88 8.47 3.28 3.85 3.35 3.95 62.50 702.90 187.30 Median 3.35 1.21 16.29 7.73 3.32 2.34 2.96 4.05 46.10 595 85 Q1 2.45 0.90 10.47 7.13 3.00 0 1.67 3.04 34.20 579 65 Min 1.02 0.01 2.13 5.21 2.56 0 1.32 2.64 15.30 108 31 Unit s s km/h m m m m m m veh/h ped/h Abbreviation TTP logTTP CarSpeed_mean Cross_length Lane_width Parking_width Sidewalk_width Cross_width BusStop_dist Veh_vol Ped_vol Variable Time-to-passage Logarithm of TTP Vehicles’ mean speed Length of the crosswalk Average width of the lanes Width of the street to park Average width of the sidewalk Width of the crosswalk Distance to a bus stop Motorized traffic volume Pedestrian traffic volume
ANALYSIS OF PEDESTRIAN BEHAVIOUR IN REAL ENVIRONMENT 40 Figure 2.10 – Boxplot of TTCmin as a function of street. Table 2.12 presents the statistical summary of the quantitative variables used in the modelling analysis of TTCmin. Regarding the pedestrians’ age, most of the sample (72.42 %) is composed of data gathered from road crossings done by adult pedestrians, i.e ., those belonging to the [20 – 40[ and [40 – 60[ ages groups. Concerning pedestrians’ sex, the sample is balanced, existing few more records of crossings performed by females (55.77 %) than male pedestrians. On the other hand, 65.36 % of all the considered pedestrian-vehicle encounters took place on a crosswalk located on a collector road. About the remaining variables, both those related to the road and pedestrian infrastructure characteristics and traffic volumes, such as for TTP analysis, the minimum and maximum values presented in Table 2.13 are in line with those previously presented in Table 2.2, Table 2.3, Table 2.4, and Table 2.5, with the exception for Q1, Q3, Median, Mean, SD, and Coefficient of variation values. Table 2.12 – Frequencies of the qualitative variables considered to model the TTCmin. Variable Abbreviation Group Absolute Frequency Relative Frequency (%) Pedestrian’s age PartAge < 20 50 10.89 [20 - 40[ 223 48.58 [40 - 60[ 114 24.84 > 60 72 15.69 Pedestrian’s sex PartGen Female 256 55.77 Male 203 44.23 Road classification Road_class Collector 300 65.36 Local 159 34.64
CHAPTER 2 41 Table 2.13 – Statistical summary of the quantitative variables considered to model the TTCmin. Coef. of Variation 0.33 0.33 0.57 0.23 0.11 0.89 0.53 0.21 0.70 0.42 0.89 SD 0.97 0.34 11.02 2.04 0.35 3.26 1.82 0.86 40.56 285.67 182.27 Max 5.71 1.74 80.57 13.80 3.87 10.12 6.57 5.35 195.20 1286 631 Q3 3.53 1.26 25.44 9.96 3.57 5.17 4.17 4.48 87.60 760 276 Mean 2.95 1.03 19.50 8.72 3.23 3.68 3.42 3.99 58.11 674.50 205.90 Median 2.86 1.05 17.87 7.85 3.26 4.23 2.96 4.05 46.10 595 85 Q1 2.27 0.82 10.89 7.13 3.00 0 1.67 3.29 22.40 528 65 Min 1.03 0.03 2.13 5.21 2.56 0 1.32 2.64 15.30 108 31 Unit s s km/h m m m m m m veh/h ped/h Abbreviation TTCmin logTTCmin CarSpeed_mean Cross_length Lane_width Parking_width Sidewalk_width Cross_width BusStop_dist Veh_vol Ped_vol Variable Time-to-passage Logarithm of TTP Vehicles’ mean speed Length of the crosswalk Average width of the lanes Width of the street to park Average width of the sidewalk Width of the crosswalk Distance to a bus stop Motorized traffic volume Pedestrian traffic volume
ANALYSIS OF PEDESTRIAN BEHAVIOUR IN REAL ENVIRONMENT 42 The statistical summary of the quantitative variables used in the modelling analysis of TTCmin presented in Table 2.13 shows that, in general, the TTCmin ranged from 1.03 to 5.71 s (m = 2.95 s; sd = 0.97 s). Such as TTP, although it is not so easy to notice, the difference between the maximum value of TTCmin (Max) and the value of the third quartile (Q3) is bigger than the difference between the value of the first quartile (Q1) and its minimum value (Min), indicating a positively skewed distribution of the TTCmin. In this way and for the same justification used for TTP, instead of considering the TTCmin as the model’s dependent variable, a new variable was created based on its natural logarithm – logTTCmin. The Pearson correlation analysis revealed that the pedestrian traffic volume was the variable with the highest correlation with TTCmin (ρ = 0.18), although it was still a low value (|ρ| < 0.30). The null model was firstly fitted (Expression 2.12), but, such as in the TTP model, the results showed that the unobserved heterogeneity of logTTCmin among the streets may not be captured using the LMM technique. logTTCmins,i = β0,i + b0,i + εs,i, (2.12) b0,i ~ N (0, σP2), εs,i~ N (0, σε2), s = 1st, …, 12th street, and i = 1st, …, 459th observation The first iteration of the process was then to fit a simple linear regression to identify the variables with a significant effect on logTTCmin. The results obtained for the first iteration of the TTCmin model are presented in Table 2.14. None of the considered variables were statistically significant in this first iteration. The iterative process continued with the removal of the Parking_width variable (p = 0.87). The 11th iteration was the last one. It only considered the Veh_vol, Ped_vol, and Cross_width variables (Expression 2.14). The results of the final iteration of the TTCmin model are presented in Table 2.15.
CHAPTER 2 43 logTTCmin,s,i = 1.0370 -0.0591 Cross_widths,i + 0.0002 Veh_vols,I + 0.0004 Ped_vols,i + εs,i, (2.13) εs,i~ N (0, σε2), i = 1st, …, 459th observation Table 2.14 – Results of the 1st iteration of the TTCmin model. β Std. Error p-value Intercept 2.1006 0.7204 < 0.01 Cross_length - 0.0269 0.0218 0.217 Lane_width - 0.1943 0.1780 0.276 Parking_width 0.0023 0.0136 0.866 Sidewalk_width 0.0110 0.0125 0.378 Cross_width - 0.0949 0.0349 < 0.01 BusStop_dist - 0.0002 0.0005 0.597 Veh_vol 0.0002 0.0001 0.242 Ped_vol 0.0006 0.0002 < 0.01 CarSpeed_mean - 0.0013 0.0016 0.424 Road_class (ref. Arterial) Local - 0.2176 0.1414 0.125 PartAge (ref. [20 – 40[) < 20 - 0.0025 0.0397 0.950 [40 – 60[ 0.0369 0.0558 0.509 > 60 0.0085 0.0465 0.855 PartGen (ref. Female) Male -0.0223 0.0323 0.489 ε 0.34 R2 0.08 Like in the TTP analysis, the low value of the coefficient of determination (R2 = 0.06) shows that the linear regression model is not the best technique to explain the variation of logTTCmin. However, this analysis aimed primarily to identify the variables with a significant effect. In this way, the increasing of the crosswalk width leads to a decreasing of the logTTCmin values (β = - 0.0591, p < 0.01). The increase of both the motorized (β = 0.0002, p < 0.01) and pedestrian traffic (β = 0.0004, p < 0.01) volumes leads to higher logTTCmin values.
ANALYSIS OF PEDESTRIAN BEHAVIOUR IN REAL ENVIRONMENT 44 Table 2.15 – Results of the final iteration of the TTCmin model. β Std. Error p-value Intercept 1.0370 0.0775 < 0.01 Cross_width -0.0591 0.0226 < 0.01 Veh_vol 0.0002 0.0001 < 0.01 Ped_vol 0.0004 0.0001 < 0.01 ε 0.33 R2 0.06 The Normal quantile-quantile (Q-Q) plot shows that there are no significant deviations from the normality assumption (Figure 2.11(a)). Any systematic increase or decrease in the variance of residuals was verified (Figure 2.11(b)). (a) (b) Figure 2.11 – Verification of the assumptions for the TTCmin model: (a) Q-Q plot; (b) Standardized residual versus fitted values. 2.4. Discussion The analysis carried out in this chapter was mainly divided into two distinct parts. The impact of variables belonging to three different groups, namely the characteristics of the built environment, the characteristics of pedestrian and motorized traffic, and the demographic characteristics of pedestrians, was first
CHAPTER 2 45 assessed in terms of pedestrian crossing decision-making through TTP and, subsequently, in terms of the severity of the encounter between pedestrians and vehicles through TTCmin on the considered crosswalks. As referred in the introductory section of this chapter, some authors argue that vehicle approaching speed has the most important role in pedestrian crossing decision-making (Granié et al. , 2014; Sucha et al. , 2017), and for others, it is the single variable appropriate to describe pedestrian safety and feeling of safety (Várhelyi, 1998). The results obtained in the analysis of TTP showed that the vehicle approaching speed has, in fact, a significant impact on pedestrian crossing decision. This is in line with Liu and Tung (2014) conclusions, which states that higher speeds can precipitate riskier crossing decisions, leading pedestrians to accept shorter time gaps and thus putting their safety and integrity at risk. Vehicle speed was not, however, the only significant variable. The transversal width of the street occupied by demarked parking places has played a role in defining TTP. According to Granié et al. (2014), the lack of parking places can induce a sensation of discomfort and lack of security in pedestrians when they decide and cross the road because it leads them to infer the existence of low volumes of pedestrian and motorized traffic, which, consequently, implies the practice of higher speeds by the drivers. Indeed, the TTP model results showed a significant positive effect of the street’s parking width, which means that the greater the width of the street used for parking places, the safer the crossing decision. The explanation for that could be based on two distinct points or on the combination of both: (i) the practice of lower speeds by drivers due to the parking, agreeing, in part, with the justification of Granié et al. (2014); or (ii) a greater level of caution and prudence felt by pedestrians in crossing judgement motivated by the parked vehicles which act as masks of the approaching vehicles. Since it was a variable with significant influence on pedestrians’ crossing decision, it was decided to evaluate which variables related to the characteristics of the road and pedestrian infrastructure and the characteristics of traffic have significant effects on the vehicles approaching speed. The results of this part of the work support the view that larger roads, with wider lanes and longer crosswalks, provide favourable conditions for the practice of higher speeds by drivers (Sucha et al. , 2017; Turner et al. , 2006; Zegeer et al. , 2006). Furthermore, they are also in line with the results of Leden (2002), which points that bigger pedestrian traffic volumes lead to lower approaching speeds of vehicles. This can also be
ANALYSIS OF PEDESTRIAN BEHAVIOUR IN REAL ENVIRONMENT 46 explained by the same justification presented by the author, who refers that the decreasing of a vehicle approaching speed is due to the greater driver alertness about the presence of pedestrians. The distance to the closest bus stop was the other variable that significantly affected vehicles’ approaching speed. The greater the distance from the crosswalk to a bus stop, the lower the vehicles’ mean approaching speed. The existence of bus stops near the crosswalk can affect the detection of pedestrians by drivers. They can be masked by buses, by other pedestrians, and even by the bus stop itself. Indeed, the results seem to indicate that it is easier for drivers to detect pedestrians approaching the crosswalk and adjust the vehicle speed when their sight view is clean. The results of the model of approaching speed have also allowed clarifying the role of the width of the street to park in pedestrians’ crossing decision-making. This variable had no significant effect on the vehicles approaching speed, which indicates that, it only affects the pedestrians’ decision-making. A possible explanation for this is the greater feeling of caution and prudence in crossing judgement task that this characteristic can promote and not a possible repercussion of the lower speeds practiced by drivers due to the parking manoeuvres, such as suggested by Granié et al. (2014). Regarding the severity of the pedestrian-vehicle encounter, the results of the model of TTCmin showed that only three of all the considered variables had a significant effect: the width of the crosswalk, the pedestrian traffic volume, and motorized traffic volume. If the crosswalk is wider, it should be easier for drivers to detect it and thus start to slow down the vehicle at a bigger distance, leading to longer TTCmin and safer crossings or encounters. However, this was not observed. The effect of the width of the crosswalk on TTCmin was in fact opposite to the expected. One explanation for this may be a greater sense of comfort and safety felt by pedestrians, leading to an excessive level of confidence during the crossing, which may jeopardize their safety. In addition to having an impact on the approaching speed of the vehicles which influenced the pedestrians’ crossing decision, the pedestrian traffic volume also has a positive effect on TTCmin. As previously explained, this effect can be explained by the increase of the detectability of pedestrians by drivers (Leden (2002). On the other hand, the effect of motorized traffic volume was in the opposite direction of conclusions of Leden (2002). In this case, the higher the motorized traffic volume, the safer the pedestrian-vehicle encounters. As well as higher pedestrian traffic volumes can increase the drivers’
CHAPTER 2 47 alertness, greater motorized traffic volumes can affect pedestrian’s caution. Besides, depending on road capacity, higher motorized traffic volumes can make high speeds impossible to be practiced by drivers. The variables related to the pedestrian characteristics, namely their age and sex, did not show an influence on pedestrian crossing safety, neither on the crossing decision making nor on the severity of conflict in the observed cases. This is in line with the conclusion of the study of Papadimitriou et al. (2016b) and contrary to what was expected and verified in other studies (Bernhoft and Carstensen, 2008; Ezzati Amini et al. , 2019; Hamed, 2001; Holland and Hill, 2007; Johansson et al. , 2004; Moyano Dıaz, 2002; Rosenbloom et al. , 2008). In general, the results of the three models show the biggest limitation of the study presented on this chapter: the observed TTP and TTCmin are not linearly explained by a simple regression. However, this part of the work aimed to identify the variables related with road and pedestrian infrastructure, pedestrians’ demographics, and pedestrian and motorized traffic characteristics on pedestrians’ decision and interaction with motorized vehicles when crossing the road. In order to assess and compare the effect of the different variables on the considered indicators, other modelling techniques should be considered taking into account the statistical characteristics of the data sample. Future work should complement this approach with data gathered in the same streets in to increase the dimension of the sample and to give more robustness to the results. It would be also interesting to consider other different streets of the same cities or even other cities in the study to assess the impact of sociocultural characteristics and city dimension on pedestrian crossing decision and safety and to increase the variability of the road and pedestrian infrastructure characteristics, allowing for the consideration of other variables which were not considered, such as the number of lanes, the land use, the road pavement, etc. 2.5. Conclusions The study presented in this chapter aimed to analyse the importance of factors linked to pedestrians’ demographics, road and pedestrian infrastructure, and traffic characteristics on pedestrian crossing safety, by assessing its influence on pedestrian crossing decision-making and the severity of pedestrian-
ANALYSIS OF PEDESTRIAN BEHAVIOUR IN REAL ENVIRONMENT 48 vehicle encounters. Video recordings were performed in twelve different crosswalks located in two different Portuguese cities, Guimarães and Braga. The trajectories and speeds of pedestrians and vehicles were collected using automated video analysis software, and, from these data, indicators such as TTP and TTCmin were calculated. The process of video recording and analysis was a long one which involved many hours of human and computational analysis. In addition, variables such as the pedestrians’ age and sex were not always easy to estimate, since, in order to significantly cover an area surrounding the crosswalk, the camera always had to be placed at a considerable distance and height. However, only in this way was it possible to extract a big and useful amount of data on pedestrians and vehicles. The models constructed to analyse the gathered data seem to support the view that factors such as speed and width of the crosswalk had a direct negative effect on pedestrian crossing safety, with higher values of the first one leading pedestrian to accept lower TTP to cross and higher values of the second one leading to the registration of minors TTCmin values. In contrast, the street’s width intended for parking places, due to the influence on the crossing decision, and the volumes of pedestrian and motorized traffic, due to the severity of the encounter, revealed a favourable effect on pedestrian safety. The lane width, the length of the crosswalk, the distance to the nearest bus stop, and again the pedestrian traffic volume significantly influenced the approaching speeds of vehicles and, thus, the safety of the most vulnerable road users, given that: the higher the first two, the higher the speeds practiced by drivers; the higher the remaining two, the lower the speeds practiced by drivers.
49 3. THE INFLUENCE OF NOISE EMITTED BY VEHICLES ON PEDESTRIAN CROSSING DECISION-MAKING 3.1. Introduction Cars are inherently noisy machines. This noise may come from its internal components, such as the engine, exhaust, and, to a lesser extent, fan and structural vibration, and from external sources, namely tire–road contact and air turbulence. The former are the primary source of noise at lower speeds, which are usually observed in urban contexts. The latter become relevant at higher speeds, more often seen on rural roads or highways. Despite several technical developments on noise attenuation technology both in and outside the vehicle, motorized vehicles are currently one of the most important sources of noise pollution (Stelling-Konczak et al. , 2015). Noise disturbs sleep, interferes in complex task performance such as school performance, modifies social behaviour, and causes emotional annoyance (Freitas et al. , 2012; Mendonça et al. , 2013; Soares et al. , 2017; Stansfeld and Matheson, 2003). This problem is well acknowledged by public authorities, vehicle manufacturers, and road industries that have continuously tried to find ways to eliminate or at least attenuate the noise emissions (Freitas et al. , 2012; Mendonça et al. , 2013; Soares et al. , 2017; Stansfeld and Matheson, 2003; Stelling-Konczak et al. , 2015). While vehicle noise is often regarded as an undesirable sub-product of transportation, it also has an important role in the interaction between vehicles and other road users, most notably, pedestrians and cyclists. Particularly in urban areas, vehicular noise often acts as a cue for vulnerable road users, improving their perception of speed and distance and calling their attention to approaching traffic. This dual nature of vehicular noise in relation to other agents raises some concerns regarding the increasing introduction of hybrid/electric vehicles that are considerably quieter than their combustion counterparts. On the one hand, hybrid and electrical vehicles emit less or no engine noise. At speeds below 30 km/h, they can produce a noise almost 4 dB(A) lower than a combustion engine-powered vehicle (Verheijen and Jabben, 2010; Wogalter et al. , 2001). Thus, generalized adoption can significantly reduce the levels of environmental noise and related subjective annoyance. However, there is the risk of an increase in the number of accidents involving vulnerable road users – which already make up for a large part of the
THE IMPACT OF THE AUDITORY CUES AND VEHICLE’S KINEMATICS ON PEDESTRIANS’ CROSSING DECISION-MAKING 56 3.2.4 Instruments The experiment was conducted in a room where the CAVE type system is located. Three DLP Christie Mirage S + 4K projectors with a resolution of 1400 × 1050 pixels are placed side by side generate an 8 m wide scene on a 9 m projection screen. They are capable of 3D stereoscopic projection, which was used with a frame rate of 60 fps. Participants, wearing 3D glasses were placed on the opposite side of the screen. The computational effort was distributed through 4 Dell Precision R7610 rack workstations, equipped with Nvidia Quadro K5000 graphics card and Intel Xeon E5 - 2600 processors. Three of them were used to render the projection itself, while the fourth was used for monitoring. BlenderVR software (Katz et al. , 2015) was used to control the simulation, paired with a VICON motion capture system. VICON reflective tracking points were placed on a set of headphones worn by participants. By knowing the position of the participants, it was possible to adjust the visual scene to their perspective, increasing the feeling of presence. The room was kept dark throughout the experiments, with the exception of the projection and the VICON’s infrared lights. Participants were placed at 2 m from the screen, rotated so that their sagittal plane formed a 35° angle with the screen. The projection viewpoint was such that the participant was placed on the sidewalk of the virtual scenario, facing the road. CPB sounds were played synchronously with the corresponding visual stimuli on the headphones, using VLC media player. The sound was amplified through a Sony TA-AV570 Audio Video Amplifier. Acoustic levels were calibrated to ensure they were equal to the ones registered during the recording sessions. 3.2.5 Experimental procedure While listening to the instructions, participants were placed in the predefined location, where they remained throughout the experiment. They were equipped with the head tracker and the 3D glasses and asked to hold a computer mouse. They were tasked with indicating, in each trial, the moment when they decided to cross the street if they felt safe to do that, clicking on any button of the mouse. They were also instructed to avoid moving or rotating their heads as much as possible during the experiment to minimize the difference between the virtual and the perceived position of the sound source (vehicle). The experimental scene was set so that the participants could see the vehicle from the start of the stimulus presentation.
CHAPTER 3 57 Participants completed an experimental session formed by two main blocks, one using the 25A scenario and others using the TP scenario. The experiment was preceded by a training block composed of 4 stimuli. Depending on the participants, there was a gap of 5 min between the two main blocks. The stimulus presentation continued after every click. However, participants were instructed to make their decision before the car stopped or passed by them, and the crossings were considered valid only for those cases. 3.2.6 Analysis The influence of the several variables addressed in this study on the participants’ crossing decisionmaking was analysed in terms of percentage of crossings, to infer about the impact of the considered variables (speed pattern and auditory condition) on the effective decision of the participants; response time, to evaluate the time that the participants needed to make their decision when they decided to cross; and TTP, which, although indirectly related to the response time, can ultimately serve as a risk-taking indicator. Here, lower TTPs at the crossing moment were assumed to be indicative of a riskier behaviour, as the participant would have less time to cross the road in a real situation. The percentage of crossings was calculated, for each participant, considering the number of answers, i.e. , the trials for which they have clicked the computer mouse before the vehicle has stopped or passed in front of them, and the total number of trials per condition. For those trials in which the participant did not click, it was assumed that the participants would only cross after the vehicle passed, and no conflict was considered. The response time, which was the time from the start of the stimulus presentation to the moment the participant clicked the mouse, was also registered. The TTP was calculated based on Expression 2.2. As mentioned in section 3.1., an impact of the noise level on the participants’ decision-making was expected. With stimuli with higher noise levels, such as those related to the approach of the gasoline vehicle (see Table 3.3), the participants would be able to better estimate the vehicle’s trajectory and would risk less, which would be translated into higher TTP values and a lower percentage of crossings. A faster decision-making for the gasoline vehicle due to the facilitated trajectory estimation was also expected. In turn and following the same increased saliency of the approaching cues, faster decisions
THE IMPACT OF THE AUDITORY CUES AND VEHICLE’S KINEMATICS ON PEDESTRIANS’ CROSSING DECISION-MAKING 58 and less risky behaviour for electric vehicles when compared with those in which no sound was presented were also expected to be found. The influence of the variables such as the auditory condition and the vehicle speed pattern on the percentage of crossings was assessed using a three-way repeated-measures ANOVA and Bonferronicorrected post-hoc tests. Due to the presence of missing values in the database regarding the trials that participants did not feel safe to cross, LMMs were used to assess the influence of the auditory condition and the vehicle speed pattern on response time and on TTP , accounting for repeated measures. 3.3. Results 3.3.1 Percentage of crossings Figure 3.3 shows the percentage of crossings for all experimental conditions. In general, participants based their decision on the movement of the vehicle, in terms of speed and distance, crossing mainly during trials where the vehicle speed decreased. Figure 3.3 – Percentage of crossings and respective mean, standard error as a function of auditory condition, per vehicle speed pattern. Although slight differences in the percentage of crossings could be observed, namely between no sound and the other two auditory conditions, it is not possible to state that the type of sound emitted by the
CHAPTER 3 59 vehicle, or its absence, has influenced the percentage of crossings. The very low percentage of crossings in the Constant speed stimuli should also be highlighted. The participants did not feel, overall, that it was safe to cross when a vehicle signals no intent of slowing down, and in the particular case of the gasoline condition, none of the participants decided to cross during in these conditions. These observations are confirmed by a two-way repeated-measures ANOVA which analyse the role of auditory condition and the vehicle speed pattern as factors affecting the crossing decision. The auditory condition did not significantly influence the participants’ percentage of crossings, F(2, 18) = 0.90, p = 0.43, while main effects were found for speed pattern, F(2, 18) = 1580.54, η2 = 0.99, p < 0.01. Bonferroni post-hoc test indicated that the percentage of crossings did not significantly differ between the Stop (m = 96.67 %; sd = 6.61 %) and Slow Down (m = 98.33 %; sd = 4.61 %) patterns, but they were significantly higher than that regarding Constant speed pattern (m = 2.67 %; sd = 7.85 %). A significant auditory condition × speed pattern interaction on participants’ percentage of crossings was also found, F(4, 36) = 2.89, η2 = 0.24, p = 0.04. However, considering the results shown in Figure 3.3, the effect of this interaction was mainly due to the great relevance of the speed pattern effect on the percentage of crossings and not exactly to that regarding the interaction between the two variables, as shown by the small effect size value (η2). Bonferroni post-hoc test indicated the percentage of crossings is only significantly different when the auditory conditions at Constant speed (electric: m = 2.00 %; sd = 4.22 %; gasoline: m = 0 %; sd = 0 %; no sound: m = 6.00 %; sd = 12.65 %), Slow Down (electric: m = 100.00%; sd = 0 %; gasoline: m = 99.00 %; sd = 3.16%; no sound: m = 96.00 %; sd = 6.99 %), and Stop (electric: m = 98.00 %; sd = 4.22 %; gasoline: m = 99.00 %; sd = 3.16 %; no sound: m = 93.00 %; sd = 9.49 %) patterns were compared. No significant differences existed between the percentage of crossings referring to the different auditory conditions verified for the Stop and the Slow Down patterns. 3.3.2 Response time Figure 3.4 shows the aggregated cumulative percentage of crossings as a function of time. It is noticeable that the few crossings with Constant speed stimuli all occurred in the initial 2 s after the beginning of the stimuli presentation. For the other speed patterns, there were considerably more crossings. Nevertheless,
THE IMPACT OF THE AUDITORY CUES AND VEHICLE’S KINEMATICS ON PEDESTRIANS’ CROSSING DECISION-MAKING 60 in both the Slow Down and Stop conditions, most responses were given after the vehicle speed began decreasing. The response time was examined with a LMM considering the vehicle speed pattern and the auditory condition variables. The time participants have taken to respond was significantly longer when the vehicle stopped (m = 3.88 s; sd = 0.86 s; p < 0.01) and when it just slowed down (m = 3.15 s; sd = 0.87 s; p < 0.01) than when it passed at constant speed (m = 1.36 s; sd = 0.41 s). The response time was significantly longer for the approaching eletric vehicle (m = 3.52 s; sd = 0.99 s) than for the gasoline combustion one (m = 3.40 s; sd = 0.79 s; p = 0.05). The no sound condition (m = 3.53 s; sd = 1.10 s; p = 0.44) did not significantly differ from the electric one. It is important to mention that existing differences between response times obtained for the gasoline vehicle and the other conditions may be partially explained by a small difference in the speeds used by the model defining movement of the cars, recorded along the passage of the vehicle when the noise was acquired. During the acquisitions, the two cars were driven by the same person to minimize the human error induced in the speed control. However, due to the different sensitivity of the vehicles’ systems and human factors of the professional driver, some differences in the order of 1.70 km/h, on average, were found in the vehicle speed. Figure 3.4 – Percentage of cumulative crossings aggregated for all participants as a function of response time.
CHAPTER 3 61 Figure 3.5 shows the number of crossings and corresponding response time per participant in each condition. Only 2 out of 10 participants felt able to cross the road when the vehicle was approaching them at 30 km/h. Besides also being the only ones to cross during the Constant speed condition, these two were also the ones who made crossing decisions more quickly in the other conditions. The results are remarkably consistent, and Figure 3.5 exhibits the low variability in the responses of all participants. Figure 3.5 – Distribution of the number of crossings with the mean standard error (horizontal lines) of the response time, by the participants, as a function of response time, per speed pattern and auditory condition. Considering all the results of the analysis of the percentage of crossings and the response time, it is noticeable that the participants felt more opportunities to cross the road safely when the vehicle speed varied, namely in stimuli where the car had the pattern of stopping and slowing down. In the responses given by the participants to the stimuli related to the Stop speed pattern, it was possible to verify some differences between the different auditory conditions. In these cases, and for the gasoline combustion vehicle, participants made their decision more quickly than in the other two types of stimuli.
THE IMPACT OF THE AUDITORY CUES AND VEHICLE’S KINEMATICS ON PEDESTRIANS’ CROSSING DECISION-MAKING 62 3.3.3 Time-to-passage The values of TTP were very similar across conditions. Nevertheless, a slightly lower value of TTP for the Stop and Constant speed conditions was noticeable when compared with the Slow Down condition. Regarding the type of auditory condition, lower values of TTP were found when the gasoline vehicle approached than in the others. The highest values of TTP were found in no sound condition (Figure 3.6). Figure 3.6 – Mean TTP and respective mean standard error as a function of auditory condition, per vehicle speed pattern. The results of the LMM confirmed what is shown in Figure 3.6. The auditory condition had a significant influence on the TTP. The obtained values of TTP regarding the no sound stimuli (m = 3.01 s; sd = 1.34 s; p = 0.02) were significantly higher than those obtained with the approaching of the electric vehicle (m = 2.80 s; sd = 0.83 s). When the gasoline combustion vehicle (m = 2.61 s; sd = 0.81 s; p = 0.04) approached the participants, the TTP was significantly lower than observed for the remaining auditory conditions. Regarding the speed pattern, the results of the model showed the TTP for the Slow Down (m = 3.00 s; sd = 0.30 s; p = 0.40) and Stop (m = 2.62 s; sd = 1.42 s; p = 0.81) patterns was not significantly different from that verified for the Constant speed pattern (m = 2.59 s; sd = 0.44 s). Because TTP is the result of dividing the vehicle distance by the approaching velocity, the role of these two components was analysed in more detail. Two LMMs were developed to assess the influence of the auditory condition on the distance at which the vehicle was and the speed it was going at the moment
CHAPTER 3 63 the participants decided to cross. The analysis in Figure 3.7 shows that these distances were very similar in all auditory conditions, considering each speed pattern separately. Figure 3.7 – Mean distance between the vehicles and participants at the moment of the participants’ responses and respective mean standard error as a function of auditory condition, per vehicle speed pattern. The results of the model show that the recorded distances at the moment of response did not differ in the no sound condition (m = 9.45 m; sd = 4.98 m; p = 0.70) or in the gasoline condition (m = 9.50 m; sd = 3.54 m; p = 0.39) when compared with those registered when approaching the electric vehicle (m = 9.37 m; sd = 4.12 m). Nevertheless, this variable appeared to be affected by the vehicle speed pattern. In the Stop pattern (m = 6.82 m; sd = 3.03 m; p < 0.01) the registered distances were shorter than those of the Slow Down pattern (m = 11.76 m; sd = 3.53 m; p < 0.01), which, in turn, were shorter than the Constant pattern (m = 18.94 m; sd = 2.78 m). Regarding the vehicle speed at the moment of the participants’ response, Figure 3.8 shows a great similarity between the three auditory conditions. However, in the results of the Stop pattern stimuli, the participants crossed with higher vehicle approaching speeds when they were presented with the sound of a gasoline combustion vehicle. The model results show that in fact the vehicle speed at the moment of response was significantly higher for gasoline stimuli (m = 13.72 km/h; sd = 4.92 km/h; p = 0.02) than for electric stimuli (m = 12.85 km/h; sd = 5.51 km/h). Non-significant differences were found between the speeds of the no sound (m = 12.47 km/h; sd = 6.34 km/h; p = 0.13) and electric condition.
THE IMPACT OF THE AUDITORY CUES AND VEHICLE’S KINEMATICS ON PEDESTRIANS’ CROSSING DECISION-MAKING 64 Figure 3.8 – Mean speed of the vehicle at the moment of the participants’ response and respective mean standard error as a function of auditory condition per vehicle speed pattern. As in the case of distance, the vehicle approaching speed at the time of the participants’ cross significantly depended on the vehicle speed pattern. In the Stop pattern (m = 11.34 km/h; sd = 5.96 km/h; p < 0.01), the registered speeds were lower than those of the Slow Down pattern (m = 14.31 km/h; sd = 4.43 km/h; p < 0.01), which, in turn, were lower than those of the Constant pattern (m = 26.37 km/h; sd = 0.93 km/h). Nevertheless, one should regard the differences between the Constant pattern and the other conditions with care, due to the lower number of observations ( i.e. , lower number of crossing decisions) in this particular condition. 3.4. Discussion On the one hand, pedestrian crossing decision-making appears to be based mainly on the readily assessable visual information of an approaching vehicle, specifically its speed and distance (Oxley et al. , 2005), on the other hand, auditory cues can play an important role in both detecting and improving locating of approaching vehicles (Barton et al. , 2013; Barton et al. , 2012). The increasing presence of hybrid and electric vehicles raises important questions about the impact of auditory cues on pedestrian safety, especially in situations of conflict between pedestrians and vehicles, such as in cases of crossing the road when a vehicle is approaching (Emerson et al. , 2013; Emerson et
CHAPTER 3 65 al. , 2011; Wiener et al. , 2006; Wogalter et al. , 2001). The purpose of this study was to contribute to this ongoing discussion by assessing the influence of the type of sound emitted by a vehicle and the auditory cues on pedestrian crossing decision-making, without neglecting the role of the vehicle speed and distance, as well as its movement pattern. Three types of stimuli were presented to the participants, corresponding to the approach of a vehicle with different movement characteristics: a vehicle emitting the sound of an electric car, a vehicle emitting the sound of a gasoline combustion car, and a vehicle with no audio component. The results show that, contrary to the hypothesis of this study, the type of emitted sound had a negligible influence on the number of times the participants decided to cross the road. On the other hand, the movement pattern of the approaching vehicle seemed to play a more relevant role. In general, participants only chose to cross when the vehicle displayed signs of slowing down. For a vehicle initial distance of 30 m and a constant speed of 30 km/h, most participants assumed that it was not safe to cross the road. The analysis of the number of crossings as a function of response times confirmed this conclusion. At Constant speed stimuli, the very few crossings occurred at an early stage of the stimulus presentation, while, in stimuli where the vehicle speed decreased, the participants waited for the approaching vehicle to reach lower speeds in order to communicate their decision to cross. The gasoline combustion vehicle seemed to lead to faster crossing decisions, particularly in the Stop condition. However, this also meant that participants crossed when the speed was still relatively high, which, counter-intuitively, resulted in lower TTP values at the time of crossing decision. The shorter response times for gasoline could, at first view, indicate a better trajectory estimation for louder vehicles. However, a difference was not found between the electric and no sound conditions. In addition, when analysing vehicle distances and speeds at the time which the responses were given, it is apparent that the participants’ decision was based primarily on the vehicle distance, which was specific for each vehicle speed pattern. For each of the three different auditory conditions, participants clicked on the computer mouse when the vehicle was always at the same distance. That distance selected by the participants was greater for higher approaching speed conditions. Expectably, the similarity in distances should have been accompanied by similarities in vehicle speed, if the vehicle speed and its evolution over time were exactly the same in the three types of auditory condition. In such cases, no difference would have been observed in the TTP values and response times when the three auditory conditions were compared. However,
THE IMPACT OF THE AUDITORY CUES AND VEHICLE’S KINEMATICS ON PEDESTRIANS’ CROSSING DECISION-MAKING 72 buildings). Two streets were used (instead of just one) to minimize the risk of biases created by uncontrolled visual elements. In order to avoid repeating the description of the scenario modelling process, information about that task can be read in section 3.2.2 of Chapter 3. Figure 4.1 and Figure 4.2 presents a comparison between the virtual scenario and the real depiction of each street. (a) (b) Figure 4.1 – Comparison between the (a) virtual and (b) real scenarios regarding the 25A Street. (a) (b) Figure 4.2 – Comparison between the (a) virtual and (b) real scenarios regarding the TP Street. 4.2.3 Stimuli The development of the audio-visual stimuli relied on real-world information collected in a two-stage procedure. In the first stage, an observational study was conducted for characterizing the typical speed
CHAPTER 4 73 patterns for vehicles approaching crosswalks. In the second stage, a test vehicle was used to replicate the identified patterns in a controlled setting while the vehicle kinematic data and resulting sounds were recorded. Next, each of these stages is described. 4.2.3.1 Observational study Firstly, the video recordings of pedestrian crossings involving approaching vehicles mentioned in section 3.2.2 of Chapter 3 were carried out in each of the selected streets. Videos were recorded at 30 fps, using a GOPRO 5 Black camera, with a resolution of 1920 × 1080 pixels. The camera was placed at the height of 5 m or more and between 15 and 25 m away from the crosswalk, depending on the conditions of each location. Each video had a duration of approximately 2 hours. A more detailed video analysis than the one carried out in the study presented in Chapter 3 was performed using the Traffic Intelligence software (Jackson et al., 2013; Saunier et al., 2010). In short, for each recorded video, Traffic Intelligence identified and mapped the movement of each pixel, frame by frame, grouping them according to the similarity of their characteristics and classified each group as an object (pedestrian or vehicle). The software provided the trajectories and speed of both vehicles and pedestrians. A total of 126 observations were registered (68 on TP street and 58 on 25A street). From the video analysis, the three most observed vehicle speed patterns in pedestrian crossing situations were identified and characterized: i) the vehicle slows down and completely stops before reaching the crosswalk (Stop); ii) the vehicle slows down before reaching the crosswalk but continues its trajectory without stopping (Slow Down); ii) the vehicle maintains its trajectory without any, or with very subtle, speed changes (Constant Speed). Descriptive statistics are presented in Table 4.1. For the Stop and Slow Down patterns, Vi and Di represent respectively the speed and distance between vehicle and pedestrians at the beginning of the observation and Vf and Df represent speed and distance at the end of the braking phase of the trajectory. For the Constant speed pattern, V represents the average speed. For each variable, mean (m), maximum (max), minimum (min), and standard deviation (sd) values are presented.
THE IMPACT OF THE AUDITORY CUES AND VEHICLE’S KINEMATICS ON PEDESTRIANS’ CROSSING DECISION-MAKING 74 Table 4.1 – Mean, maximum, minimum, and standard deviation of vehicle speed and distance to pedestrian for each speed pattern. Stop Slow Down Constant Speed Vi (km/h) Di (m) Vf (km/h) Df (m) Vi (km/h) Di (m) Vf (km/h) Df (m) V (km/h) m 17.30 15.80 0 5.71 21.34 16.72 7.24 7.01 25.91 sd 8.15 5.62 0 2.40 8.46 6.21 6.56 4.14 7.87 max 34.41 27.37 0 14.06 37.59 28.20 28.39 22.04 43.18 min 3.81 6.58 0 1.26 4.43 2.94 1.07 2.69 14.27 The characteristics of the vehicle movement for the simulation (Table 4.2) were defined, taking into account the descriptive statistics presented in Table 4.1. The goal was to have a limited set of experimental conditions while still having representative values of Vi, Di, Vf, and Df. Table 4.2 – Characteristics of vehicle movement in the different conditions presented on the experiment. Condition Vi (km/h) Vf (km/h) Di. mov (m) Di (m) Df (m) 1 20 20 35 - - 2 30 30 35 - - 3 20 20 30 - - 4 30 30 30 - - 5 20 20 25 - - 6 30 30 25 - - 7 30 10 30 25 5 8 20 10 30 15 10 9 20 0 30 15 5.50 10 30 0 30 20 5.50 Perhaps because pedestrians found it unsafe to cross, almost no crossings were observed for the selected distances and speeds above 30 km/h during the observational study. So, the conditions of the study were limited to speeds equal to or lower than this value. To decrease the likelihood of participants memorizing the vehicle’s motion profiles and taking into account that in the Slow Down and Stop patterns the speed
CHAPTER 4 75 varied along the time of approaching, hampering some learning effect that could occur, the initial distance at which the vehicle appeared (Di, mov) in the Constant speed stimuli also varied. 4.2.3.2 Auditory stimuli and trajectory acquisition Following the approach used in the study described on Chapter 3, controlled binaural sound recordings of real vehicles were carried out in a closed urban road. The vehicles’ noise was recorded using a Brüel & Kjaer Pulse Analyzer type 3560-C and a Brüel & Kjaer Head and Torso Simulator (HATS) Type 4128-C equipped with Ear Simulators Type 4158-C and 4159-C. The Controlled Pass-By (CPB) movement of a Kia Ceed SW with a gasoline combustion engine and equipped with ContiEcoContact3 195/65-R15 tires was used for the recordings. CPB measurements include all vehicle noise sources, the effect of propagation mechanisms, and noise from the surrounding environment (Freitas et al. , 2012). For this reason, the recordings were performed during the night-time (20:00h – 24:00h) in a quiet zone to avoid traffic noise. To minimize the meteorological bias, all recording sessions were performed with dry pavement, wind speed below 5 m/s, atmospheric temperature between 5°C and 30°C, and pavement temperature between 5°C and 50°C as recommended in ISO 11819-1 (1997). Moreover, to represent the average conditions of the Portuguese urban roads, the sound recordings were performed in an asphalt mix (AC14) pavement with good maintenance conditions. During the recordings, the HATS was placed with its head turned 35º from the road direction, on the side of the road, at 1.55 m from its centre, and 1.66 m height (Portuguese population average height). To generate the sound samples for each of the three types of vehicle motion pattern, a driver performed the trajectories defined by the parameters in Table 4.2. The real speed, time, and distance data from the vehicle were registered at 1/8 Hz rate simultaneously with the sound recordings. The sound and vehicle position data were later synchronized, calibrated, and implemented in a virtual environment. The visual model of the vehicle used in the experiment was the same used in sound recordings.
THE IMPACT OF THE AUDITORY CUES AND VEHICLE’S KINEMATICS ON PEDESTRIANS’ CROSSING DECISION-MAKING 76 4.2.4 Instruments The experiment was conducted in the same room and using the same CAVE type system used in the study described in the previous chapter. For more information about the experimental setup and instruments used in this study, see section 3.2.4 of Chapter 3. 4.2.5 Experimental procedure Before the beginning of the experiment, each participant was placed in a predefined room point (2 m from the projection screen) where they had to stay throughout the experiment (Figure 4.3). This point corresponded, in the virtual space, to the intersection of the perpendicular to the direction of the vehicle movement and crosswalk’s axis of symmetry and the curb. Figure 4.3 – Spatial layout of the room and participant position. Participants were asked to put on the headset and the 3D glasses while listening to the instructions. They were then tasked with indicating, in each trial, the moment they decide to cross the street, clicking on the buttons of the mouse. They were also told not to press any button if they decided not to cross. Figure 4.4 shows a depiction of the experiment’s performance.
CHAPTER 4 77 Figure 4.4 – Participants’ view during the performance of the experiment. Participants completed an experimental session composed of two main blocks, one in each street scenario. The experiment was preceded by a training block composed by four crossing trials. Each one of the ten speed patterns presented in Table 4.2 was randomly repeated five times for each type of auditory condition (gasoline combustion engine – audio-visual condition and those without auditory cue – no sound condition ). Participants went through 200 trials (10 movement conditions × 5 repetitions × 2 auditory conditions × 2 streets). The two main blocks were also split into two parts so that participants could rest between each part and block for as long as they needed. The experiment lasted 1 hour. 4.2.6 Analysis As in Chapter 3, the influence of the variables addressed in this study on the participants’ crossing decision-making was also analysed in terms of the percentage of crossings, response time, and TTP (see all the definitions in section 3.2.6 of Chapter 3). The influence of the variables auditory condition , vehicle speed pattern , vehicle initial speed , and vehicle initial distance on the percentage of crossings was assessed using a three-way repeated-measures ANOVAs. The influence on response time and TTP was assessed by fitting mixed-effects regression models with random effects included for the participant and fixed effects defined for the three variables mentioned above. The use of Linear Mixed Models is justified by the existence of missing values in the data, corresponding to trials in which participants did not cross. The data analysis was done in two distinct stages. In the first stage, the conditions characterized by constant speed patterns were analysed. In the second stage, a comparison was made between the different speed patterns.
THE IMPACT OF THE AUDITORY CUES AND VEHICLE’S KINEMATICS ON PEDESTRIANS’ CROSSING DECISION-MAKING 78 4.3. Results 4.3.1 Constant speed pattern Figure 4.5 shows the crossings’ data for each participant and condition (location of the circle along the axis represents the mean response time, the size of the circle represents the number of crossings, and the line shows the standard error of the mean; participants are vertically ordered according to mean Response Time). Figure 4.5 – Participants’ responses along time of stimulus’ presentation, per auditory condition and initial speed (Constant speed pattern).
CHAPTER 4 79 It is possible to note a group of participants that frequently decided to cross, irrespective of the condition. These participants also made consistently faster decisions. The summary of descriptive statistics in terms of percentage of crossings, response time and TTP, regarding the analysis of the Constant speed pattern is presented in Table 4.3. Table 4.3 – Descriptive statistics of the percentage of crossings, response time, and TTP for each trial and vehicle Constant speed pattern only. Crossings Response time TTP Audio Cond. Vi Di.mov Mean SD SE Mean SD SE Mean SD SE (km/h) (m) (%) (%) (%) (s) (s) (s) (s) (s) (s) No sound 20 25 27.30 30.80 5.63 1.28 0.56 0.06 3.43 0.54 0.06 30 45.00 40.20 7.35 1.33 0.85 0.07 4.41 0.77 0.07 35 53.00 40.40 7.37 1.37 0.89 0.07 5.19 0.81 0.06 30 25 5.00 17.80 3.24 0.59 0.18 0.05 2.53 0.19 0.05 30 9.33 22.70 4.15 0.81 0.39 0.07 2.95 0.31 0.06 35 14.30 23.60 4.31 1.03 0.73 0.11 3.34 0.62 0.10 Sound 20 25 23.00 31.90 5.82 1.07 0.45 0.05 3.62 0.41 0.05 30 41.30 42.30 7.73 1.18 0.58 0.05 4.53 0.52 0.05 35 49.30 40.70 7.43 1.31 0.82 0.07 5.22 0.75 0.06 30 25 4.33 16.80 3.06 0.69 0.19 0.05 2.41 0.23 0.06 30 7.67 17.70 3.24 0.87 0.49 0.10 2.91 0.43 0.09 35 12.70 22.90 4.18 1.05 0.50 0.08 3.32 0.43 0.07 4.3.1.1 Percentage of crossings The percentage of crossings was examined using a three-way repeated-measures ANOVA with vehicle initial distance (3), vehicle speed (2) and the auditory condition (2) as factors. The auditory condition did not significantly influence the participants’ percentage of crossings F(1, 29) = 2.71, p = 0.11, η2 = 0.09. There was a main effect of vehicle speed, F(1, 29) = 39.35, p < 0.01, η2 = 0.58, with higher values for the 20 km/h condition compared to when it was 30 km/h. There was also a main effect of vehicle initial distance, F(2, 58) = 31.18, p < 0.01, η2 = 0.52, with Bonferroni post hoc tests showing that crossing percentages increased significantly with the initial distance. A significant speed × initial distance interaction was also found, F(2, 58) = 9.66, p < 0.01, η2 = 0.25. Bonferroni post hoc tests were conducted
THE IMPACT OF THE AUDITORY CUES AND VEHICLE’S KINEMATICS ON PEDESTRIANS’ CROSSING DECISION-MAKING 80 comparing initial distances within each level of speed. At 20 km/h, there were significant differences between all levels of distance (25 m / 30 m: p < 0.001, 25 m / 35 m: p < 0.001, 30 m / 35 m: p = 0.01). At 30 km/h significant differences were only found between 25 and 35 m (p < 0.001). In general, participants crossed more when vehicle speed was lower and its initial position was farther from the crosswalk. However, at 20 km/h, the distance increasing resulted in greater growth of the percentage of crossings (see Figure 4.6). Figure 4.6 – Percentage of crossings and respective mean standard error as a function of auditory condition, per initial distance and initial speed (Constant speed pattern). 4.3.1.2 Response time Initial analysis of response time data showed a skewness pattern typical of response times. A linear mixed model of response time with random effects included for participant and fixed effects defined for the initial speed , speed pattern, and auditory condition was fitted. Visual inspection of the residual plots showed deviations from homoscedasticity and skewness, so the model was refitted applying a logarithmic transformation to response times, which corrected the deviations. Satterthwaite's tests showed significant effects of vehicle initial speed, F(1, 838.30) = 19.00, p < 0.001, and vehicle initial distance, F(1, 837.22) = 6.01, p < 0.05, and an interaction between initial speed and initial distance, F(1, 837.38) = 3.10, p = 0.04. There was no effect of the auditory condition.
CHAPTER 4 81 The model was refitted after discarding the auditory condition, and contrasts were used to analyse the interaction between vehicle initial speed and initial distance. Differences were found, for the initial speed of 30 km/h, between initial distances of 25 and 35 m, b = 0.20, t(842.84) = 2.45, p = 0.01. No significant differences were found for the speed of 20 km/h. Considering these results and Figure 4.7 that shows the response time as a function of vehicle initial distance, initial speed and auditory condition, it is apparent that participants’ decisions to cross tended to be faster for higher speeds and when the vehicles were closer. Figure 4.7 – Mean response time and respective standard error as a function of auditory condition, per initial distance and initial speed (Constant speed pattern). 4.3.1.3 Time-to-passage Figure 4.8 shows the mean values of TTP at the crossing instant as a function of the TTP at the start of the trial. The observed TTP appears to vary linearly with the initial TTP. This was expected given that response time variations were small compared with the variation of initial TTP. The response time analysis done in the previous subsection showed that for those participants that crossed with the approaching vehicle at 30 km/h there was an attempt to compensate for the shorter available time by initiating the crossing earlier. However, TTPs were still lower than the ones observed at 20 km/h, meaning that participants who crossed at higher speeds were, in effect, taking riskier decisions.
THE IMPACT OF THE AUDITORY CUES AND VEHICLE’S KINEMATICS ON PEDESTRIANS’ CROSSING DECISION-MAKING 88 pattern stimuli, the others only took place after the vehicle began to slow down. Figure 4.13(a) shows the histogram of the number of crossings as a function of response time. One can see that most crossings concentrated around these two moments. To verify if participants were individually consistent in the moments they choose to cross, the sample of thirty participants was split into two groups, according with the individual mean response times (Figure 4.13(b)). The value of 2.20 s was used as a threshold, since it corresponds to the local minima of the complete histogram that separates the early from the late crossings. One can see that response times above this threshold generally belong to the same participants, the ones who consistently crossed in the Slow Down and Stop conditions when the vehicle speed was decreasing (Group A). Response times below threshold belong to the other group (Group B) which includes those who tended to cross in the initial moments of the presentation of the stimuli, both at constant speed (especially at 20 km/h) and in the other conditions. The first group includes a major part of participants. A Mann-Withney confirms the observation, showing that the mean values of the response time of the two groups are statistically different (W = 945808, p-value < 0.01). (a) (b) Figure 4.13 – Histogram of number of crossings by response time: (a) for the general data; and (b) by speed pattern and group of participants.
CHAPTER 4 89 4.3.2.3 Time-to-passage A model was also fitted for TTP with random effect included for participant and fixed effects defined for the vehicle initial speed , speed pattern, and auditory condition . Visual inspection of the residual plots showed deviations from homoscedasticity and skewness. So, the model was refitted after applying a log transformation to TTP values, which minimized the deviations. As TTP is directly dependent on the response time, similar effects to the ones found on the response time model were expected to exist. Satterthwaite's tests showed significant effects of initial speed, F(1,2152.80) = 61.11, p < 0.001, speed pattern, F(2,2157.30) = 3.05, p = 0.04, and an interaction effect between initial speed and speed pattern, F(2, 2151.20) = 39.27, p < 0.001. There was no effect of the auditory condition. The model was refitted discarding the auditory factor and contrasts were used to compare the different speed patterns. A first contrast compared the Constant and Slow Down speed patterns, and nonsignificant differences were found between them. A second contrast compared the Constant and Stop patterns, and it showed a significant difference between them, b = - 0.18, t(2163.03) = - 5.07, p < 0.001. A third contrast compared the Constant and Slow Down patterns for each initial speed. Significant differences were found, b = 0.22, t(2158.82) = 2.91, p < 0.01, showing that, for 20 km/h, the TTPs on the Slow Down pattern are significantly lower than in the Constant one, but for 30 km/h, the opposite is true. A fourth contrast showed the same tendency when comparing the Constant and Stop patterns for each initial speed. A significant difference was found, b = 0.52, t(2158.61) = 6.86, p < 0.001, showing that, at 20 km/h, the TTP was lower for the Stop compared with the Slow Down, but the opposite was true for the 30 km/h. It is worth noting that, although the differences are significant, the TTPs for the Slow Down patterns are close to the ones for the Constant. Nevertheless, the percentage of crossing was remarkably higher in this condition, showing that participants have either made a substantially different risk assessment or were in the belief that, as the vehicle was slowing down it would eventually stop, which in fact did not happen. The large increase in TTP for the 30 km/h is shown in Figure 4.14. At 30 km/h, most crossings happened later in the trial, when the vehicle was already at a lower speed, making the TTP higher. No significant differences were found for the auditory condition. However, Figure 4.14 shows that participants accepted higher TTPs when no sound was presented, particularly for the Stop pattern.
THE IMPACT OF THE AUDITORY CUES AND VEHICLE’S KINEMATICS ON PEDESTRIANS’ CROSSING DECISION-MAKING 90 Figure 4.14 – Mean TTP and respective standard error as a function of auditory condition, per initial speed and speed pattern. 4.4. Discussion For a vehicle approaching at constant speed, the results suggest that both speed and distance affect the crossing decisions. Crossing percentages increased with distance but were also substantially lower at the highest speed. While the experimental design used in this study does not allow a direct comparison of crossing decisions as a function of TTP with different speeds, the results do hint to an important role of speed in crossing decisions instead of a decision criterion based mostly on distance. Considering that the study population was composed of young adults, with ages between 20 and 40 years old, this agrees with past research (e.g. (Liu and Tung, 2014; Lobjois and Cavallo, 2007, 2009; Oxley et al. , 2005)) which indicates that young people are better at estimating the available time to cross and thus more apt to balance the risk of increased speed. However, these results contrast with the study of Feldstein and Dyszak (2020). In their experiment, a group of young participants seemed to use distance-based criterium in a virtual environment in contrast to a time-based criterium in real environment. The difference might be attributed to different experimental speeds. It is known that lower speeds are easier to distinguish. The virtual vehicle used by Feldstein and Dyszak (2020) moved at speeds ranging from 30 to 40 km/h. In this study, speeds between 20 and 30 km/h were used, which may have made the lower speed more salient in comparison to the higher, fostering a weighted crossing decision. The response times also showed dependency on both speed and distance. Higher speeds and shorter distances seem to have prompted faster decisions, although this result was only significant for the higher
CHAPTER 4 91 speed. This observation was not entirely unexpected. Lobjois and Cavallo (2009) compared response times for crossing decisions in a simulated environment, with and without a time constraint. They verified that participants were faster to respond under the time-constrained condition, which also favoured a distance-based decision. It is possible that, in this study, the higher speeds and close distances have prompted a similar sense of urgency, which is known to accelerate decision processes, although at the cost of accuracy in judgments (Soares et al. , 2020). This agrees with Beggiato et al. (2018), who analysed the effect of daytime, approaching vehicle speed and pedestrian’s age on the time gaps accepted to cross the road, and found that the participants took more risky crossing decisions, accepting lower time gaps with the increasing vehicle speed. An important consideration can be made regarding the individual subjects’ behaviour. It is noticeable that most participants were consistent in terms of response time and crossings percentages but differed substantially among themselves. Participants with higher crossing percentages were also the ones with shorter response times. Figure 5 shows that participants who decided not to cross at 30 km/h crossed fewer times and did so later at 20 km/h. In contrast, participants who decided to cross often at 30 km/h also decided to cross early for both approaching speeds, pointing to a more impulsive behaviour. However, the faster responses were not enough to substantially increase the TTP, which means that in a real situation, these participants would had put themselves at greater risk. Concerning the influence of the auditory condition on participants’ crossing decision, despite a slight superiority in the crossing percentages obtained in the presentation of the merely visual stimuli, mainly for the speed of 20 km/h, the applied statistical tests showed that differences were not significant. The analysis of the response time and, consequently, TTP made it even more evident that the auditory condition had no effect on the participants’ crossing decision-making. Regarding the comparison between different types of speed pattern, results also showed a significant influence of the vehicle’s initial speed on the percentage of crossings. In this case, the participants crossed more often when the initial speed of the vehicle was 20 km/h. The analysis of the results showed even more clearly a division between those participants who crossed shortly after the start of the stimulus and those who took more time to decide. As previously referred, participants with shorter reaction times tended to cross more often, showing that they are more likely to make dangerous decisions consistently. Most of the participants took a more cautious approach, with few crossings when the vehicle was
THE IMPACT OF THE AUDITORY CUES AND VEHICLE’S KINEMATICS ON PEDESTRIANS’ CROSSING DECISION-MAKING 92 approaching at a constant speed, particularly at 30 km/h. When the vehicle slowed down, they started to cross when the speed started decreasing. The longest response times in the Stop and Slow Down conditions confirm this behaviour. For both conditions, participants seem to have assumed that the deceleration meant the vehicle would stop, or at least provide them enough time to cross, irrespective of the initial speed. An indication of this can be found in the differences in the percentage of crossings between initial speeds. In the Constant speed condition, this difference is substantial with more crossings at 20 km/h. In the Slow Down and Stop conditions there is no significant difference between initial speeds. Also, crossing percentages are overall higher than in the Constant condition. The TTP analysis also seems to support this. At the initial speed of 30 km/h in the slow down condition, the TTP was remarkably lower than in the stop condition, although neither response time nor crossing percentages differed significantly. This again seems to indicate that participants made their decisions based on the perceived deceleration and not on a TTP estimation. The results highlight the role of perceived vehicle kinematics as a communication tool between vehicle and pedestrian. The early deceleration seems to have been taken by the participants as an indication that they could cross, with the initial speed and actual driver intention (stop or simply slow down) playing a less important role in the decision. This conclusion is in line with the results of recent studies that explore the role of vehicle movement as a mean of communication and coordination between drivers and pedestrians. Deceleration is normally interpreted by pedestrians as an indication that the driver has seen them and will yield the passage (Ackermann et al. , 2018; Dey et al. , 2019; Mahadevan et al. , 2018; Schmidt and Färber, 2009; Várhelyi, 1998). Drivers, in turn, may deliberately use anticipated braking as a way to signal their yielding intention, encouraging the pedestrian to cross with the vehicle still moving, speeding up the encounter and eventually preventing the need for a full stop (Risto et al. , 2017). This result has implications in the development of communication strategies between automated vehicles and pedestrians, a topic that has been receiving growing attention as the presence of driverless vehicles in our roads seems to be an approaching reality (Schneemann and Gohl, 2016; Sucha et al. , 2017). On the one hand, vehicle developers should keep in mind that speed adjustments may convey false cues regarding vehicle behaviour. On the other hand, kinematics may be a simple way to convey intention to
CHAPTER 4 93 pedestrians, although it should also be considered that judgements of available crossing time may be inaccurate and lead to risky situations (Dietrich et al. , 2020). At last, regarding the audio-visual and the merely visual stimuli, participants relied mostly on the visual information they received from the approaching vehicle to estimate the available crossing time, contradicting Barton et al. (2012) and agreeing with the conclusions of Pugliese et al. (2020) and Soares et al. (2020). These results prove that the absence of the audio component regarding the road traffic does not compromise the results obtained in pedestrian safety studies performed in a virtual environment. Table 4.5 summarizes the main findings of the present work. As a limitation, it should be underlined that this study only considers the particular situation in which the vehicle approached the pedestrian from a clearly visible position and maintaining a straight trajectory without obstacles to the participants’ view, such as in other studies developed in this research area (e.g. (Cavallo et al. , 2019; Cavallo et al. , 2009; Charron et al. , 2012; de Clercq et al. , 2019; Dommes et al. , 2012; Feldstein et al. , 2016; Simpson et al. , 2003; Zito et al. , 2015)). In cases in which the visibility of participants could be impaired, greater importance of the auditory cues in their decision would be expected, as shown by Barton et al. (2012). Another limitation of this study was the participants’ experimental task. In the experiments, participants signalled their decision by clicking on a button of a computer mouse while remaining still in a predetermined position. This approach has been used in most pedestrian simulator studies (Charron et al., 2012; Meir et al., 2015; Schwebel et al., 2008; Thomson et al., 2005; Zito et al., 2015). Some of the most recently developed simulators already allow participants to freely walk the virtual crossroad (Cavallo et al. , 2019; Deb et al. , 2018b; Feldstein et al. , 2016; Morrongiello et al. , 2015; Schneider et al. , 2021; Simpson et al. , 2003; Sween et al. , 2017). Being able to move freely turns the simulator and the experimental task more realistic and immersive. It provides participants with a complete sense of the space and their own speed and, consequently, of the time they need to initiate movement and cross the road. However, one of the objectives of this study was to analyse the role of auditory cues in the pedestrians’ crossing decision and modelling vehicular sound in a spatially congruent manner for a moving listener is far from trivial. For the sake of simplicity, it was decided to consider an experimental task where the participant chose between go/no go options by clicking on a simple button. This also allowed for lighter and less time-consuming trials for the participants.
THE IMPACT OF THE AUDITORY CUES AND VEHICLE’S KINEMATICS ON PEDESTRIANS’ CROSSING DECISION-MAKING 94 Table 4.5 – Summary of the main findings. Main Findings Agreeing with Disagreeing with When a vehicle is approaching the crosswalk at constant speed, both speed and distance affect the pedestrian crossing decisions. Oxley et al. (2005); Lobjois and Cavallo (2007); Lobjois and Cavallo (2009); Liu and Tung (2014) Feldstein and Dyszak (2020) Vehicle’s higher speeds and shorter distances lead to faster and unsafe crossing decisions. Lobjois and Cavallo (2009); Beggiato et al. (2018); Soares et al. (2020) - Assuming full visibility of the approaching vehicle, the auditory condition does not influence the pedestrians’ crossing decisionmaking. Pugliese et al. (2020); Soares et al. (2020) Barton et al. (2012) Vehicle kinematics is a relevant communication tool between vehicle and pedestrian. Vehicle deceleration is interpreted by pedestrians as an indication of the intention to yield the passage. Várhelyi (1998); Schmidt and Färber (2009); Ackermann et al. (2018); Mahadevan et al. (2018); Dey et al. (2019) - Future work may pass by the analysis of other types of trajectories (such as turning movements at intersections) and obstacles to the participants' vision (such as parked vehicles, trees, and urban furniture), to assess the general effect of the auditory cues on the pedestrian crossing decision-making in a more comprehensive way.
CHAPTER 4 95 4.5. Conclusions The main goal of this research was to analyse the role of visual and auditory cues in crossing decisions, considering different initial speeds and distances of the vehicle as well as deceleration profiles. The results support the general conclusion that the speed and initial distance of the vehicle and its speed profile impact the crossing behaviour. Contrarily, the auditory input has no major role in modulating decisions, at least when pedestrians are crossing a virtual road having perfect visibility conditions of the approaching vehicle. Regarding the vehicle motion, most participants made a decision based on the vehicle’s perceived kinematics, with the deceleration being interpreted as an indication that the vehicle would yield the passage. However, a small group of participants seems to have responded more hastily, taking less time to decide, crossing more often, including when the vehicle approached at the highest speed and from closer distances and irrespective of whether it would yield or not. These conclusions highlight the role of vehicle kinematics as an important mean of communication between vehicles and pedestrians, which should be considered in the development of new strategies to mitigate the severity of conflicts between vulnerable road users and motorized traffic. The results are also relevant for autonomous driving developers, showing that vehicle movement can be explicitly used to communicate with pedestrians and that care should be taken to prevent unintentional, misleading signals.
96 5. COMPARISON BETWEEN TWO TYPES OF EXPERIMENTAL APPROACH TO ASSESS THE PEDESTRIAN BEHAVIOUR IN VIRTUAL ENVIRONMENT 5.1. Introduction As previously referred, understanding which variables may influence the pedestrian behaviour and decision-making during conflicts with the motorized traffic, such as crossing the road, has been the goal of several studies over last years ( e . g . (Bernhoft and Carstensen, 2008; Ewing and Dumbaugh, 2009; Ezzati Amini et al. , 2019; Granié et al. , 2014; Hamed, 2001; Holland and Hill, 2007; Ishaque and Noland, 2008; Johansson et al. , 2004; LaScala et al. , 2000; Leden, 2002; Lin et al. , 2015; Moyano Dıaz, 2002; Oxley et al. , 2005; Papadimitriou et al. , 2016a, 2016b, 2017; Papadimitriou et al. , 2012; Rosenbloom et al. , 2008; Sucha et al. , 2017; Sueur et al. , 2013; Turner et al. , 2006; Zegeer et al. , 2006)). As mentioned in Chapter 1 and according to Papadimitriou et al. (2016b), Feng et al. (2021), and Deb et al. (2018a), methods for analysing pedestrian behaviour are based on field observation, survey, semicontrolled experiments, and simulation. The most common way of gathering data about pedestrians’ crossing behaviour is through video recordings of those same pedestrians (Lassarre et al. , 2012). Those recordings, however, are limited to the used camera’s field of vision and may fail to capture important parts of the interaction. Other alternatives, such as following the trajectory of pedestrians through GPS instruments or Bluetooth/Wi-Fi sensors, have also limitations, such as problems with precise location and unavailability of information regarding traffic conditions (Feng et al. , 2021; Papadimitriou et al. , 2016a). Surveys are the most often used method to obtain data for qualitative analysis. They are done through written documents, online questionnaires, face-to-face or telephone interviews (Deb et al. , 2018a). However, participants’ answers may not portray their actions in real situations, and a big sample of participants is demanded (Feng et al. , 2021).
CHAPTER 5 97 Semi-controlled experiments are usually applied to analyse factors such as gait parameters and pedestrian spatial organization along predefined paths (Cao et al. , 2018; Fu et al. , 2019; Wei et al. , 2015). However, this method has the same limitation as all pedestrian controlled experimental studies. The participants’ behaviour can be influenced by the fact they know they are being observed and analysed. Alternatively, some experiments have been performed using virtual reality simulators in which the test participant visualizes a crossing situation and must choose between go/no got options throughout clicking a simple button (Charron et al. , 2012; Meir et al. , 2015; Schwebel et al. , 2008; Thomson et al. , 2005; Zito et al. , 2015) or having a free walk on a virtual crosswalk (Cavallo et al. , 2019; Deb et al. , 2018b; Feldstein et al. , 2016; Morrongiello et al. , 2015; Simpson et al. , 2003; Sween et al. , 2017). Despite having disadvantages such as the greater need for space to carry out the experiments and dependence on expensive equipment, simulator-based experiments for studying road agents’ behaviour have several advantages compared with similar experiments conducted in real-world. They avoid most of the hurdles required to ensure participants’ safety in real environments while allowing more control over experimental conditions and tasks (Deb et al. , 2017). Pedestrian crossing simulators can be divided into simulators that rely on head-mounted displays (HMDs) or simulators that use Cave Automatic Virtual Environments (CAVE) technology (Cavallo et al. , 2019). Compared to HMD solutions, projection-based simulators allow for greater freedom of movement for the participants. By using a power-wall configuration and a motion tracking system to project the intended scenario with a perspective adjusted to the physical location of the participant, this type of simulators allows participants to conduct the act of crossing on their own, without the use of instruments such as treadmills or joysticks (Cavallo et al. , 2019). Pedestrians’ behaviour must be studied with enhanced tools to provide a complete and reliable tool for road safety managers. According to Feldstein et al. (2016), the quality of each simulator is associated with the capacity of inducing on the participants the feeling of being present in the virtual environment and not just perceiving it as a digital image, which in turn depends on the realism of the environment and the usability of the simulator, supported by the quality of the graphical representation, sound, and interaction possibilities.
COMPARISON BETWEEN TWO TYPES OF EXPERIMENTAL APPROACH TO ASSESS THE PEDESTRIAN BEHAVIOUR IN VIRTUAL ENVIRONMENT 104 Figure 5.3 – Spatial layout of the room in the dynamic approach. Figure 5.4 – Participant performing the dynamic experiment. As in the other experiments, participants completed an experimental session made of two main blocks, one using the 25A scenario and others using the TP scenario, preceded by a training block composed of 4 stimuli. There was a gap of 5 minutes or more between the two main blocks to rest, depending on the participant’s wishes.
CHAPTER 5 105 5.2.6 Analysis The influence of the several variables addressed in this study on the participants’ crossing decisionmaking was analysed in terms of the percentage of crossings, crossing start time, and TTP. Such as in Chapter 3 and Chapter 4, the percentage of crossings was calculated, for each participant, by assuming that: (i) in static approach, a decision to cross was considered in the trials in which the computer mouse was clicked before the vehicle had stopped or passed by the participants’ position; and, (ii) in dynamic approach, a decision to cross was considered in the trials in which the participants had crossed the half-length of the semi-virtual crosswalk before the vehicle had stopped or passed in front of them. Since the participants’ task was to effectively cross the virtual road, in the dynamic approach (contrary to the static approach) it was possible to count the number of crashes. The crossing start time, corresponding to the time from the beginning of the stimulus presentation until the moment when the participant clicked the mouse or took the first step on the crosswalk, was also registered. The influence of the variables experimental approach , vehicle speed pattern , vehicle initial speed , and vehicle initial distance on the percentage of crossings was assessed using a three-way repeated measures ANOVA. The crossing start time and TTP were assessed by fitting mixed-effects regression models with random effects included for the participant and fixed effects defined for the three variables mentioned above. The use of LMMs is justified by the existence of missing values in the data, corresponding to trials in which participants did not cross. The analysis of the results was done in two stages: in the first stage, the conditions characterized by Constant speed patterns are analysed, since, only in these cases, there was a variation of the vehicle initial distance; In the second stage, the data regarding all speed patterns is analysed considering only the conditions with 30 m of initial distance for the Constant speed trials. In this way, it was ensured that the analysis of the vehicles’ speed patterns effect on the participants’ responses was carried out under equal conditions. Complementarily, a comparison between the TTP obtained with the two experimental approaches’ performance and the video recordings concerning the 25A and TP streets is presented.
COMPARISON BETWEEN TWO TYPES OF EXPERIMENTAL APPROACH TO ASSESS THE PEDESTRIAN BEHAVIOUR IN VIRTUAL ENVIRONMENT 106 5.3. Results 5.3.1 Constant speed pattern Figure 5.5 and Figure 5.6 show the general distribution of the data by participant and condition (similarly to those of previous chapters, location of the circle along the axis represents the mean response time, the size of the circle represents the number of crossings and the line shows the standard error of the mean; participants are vertically ordered according to mean crossing start time). In general, with few exceptions, the participants took more time to start to cross in the dynamic approach than in static, irrespective of the condition. Table 5.4 and Table 5.5 show the summary of descriptive statistics by condition, in terms of percentage of crossings, percentage of crashes, crossing start time, and TTP, to complement the information presented on the following figures. (a) (b) Figure 5.5 – Participants’ responses along time of stimulus’ presentation, per initial distance, regarding the 20 km/h speed, for Constant speed pattern and for each experimental approach: (a) dynamic; and (b) static.
CHAPTER 5 107 (a) (b) Figure 5.6 – Participants’ responses as a function of the time of stimulus’ presentation, per initial distance, regarding the 30 km/h speed, for Constant speed pattern and for each experimental approach: (a) dynamic; and (b) static. Table 5.4 – Descriptive statistics of percentage of crossings and crashes for each trial regarding the Constant speed pattern. Crossings Crashes Experimental Approach Vi Di.mov Mean SD SE Percentage (km/h) (m) (%) (%) (%) (%) Static 20 25 23.00 31.90 5.82 - 30 41.30 42.30 7.73 - 35 49.30 40.70 7.43 - 30 25 4.33 16.80 3.06 - 30 7.67 17.70 3.24 - 35 12.70 22.90 4.18 - Dynamic 20 25 12.30 19.80 3.61 59.50 30 54.30 39.90 7.28 0.61 35 74.00 35.00 6.39 0.90 30 25 0 0 0 0 30 1.33 7.30 1.33 75.00 35 5.00 16.80 3.06 20.00
COMPARISON BETWEEN TWO TYPES OF EXPERIMENTAL APPROACH TO ASSESS THE PEDESTRIAN BEHAVIOUR IN VIRTUAL ENVIRONMENT 108 Table 5.5 – Descriptive statistics of crossing start time and TTP for each trial regarding the Constant speed pattern. Crossing start time TTP Experimental Approach Vi Di.mov Mean SD SE Mean SD SE (km/h) (m) (s) (s) (s) (s) (s) (s) Static 20 25 1.07 0.45 0.05 3.62 0.41 0.05 30 1.18 0.58 0.05 4.53 0.52 0.05 35 1.31 0.82 0.07 5.22 0.75 0.06 30 25 0.69 0.19 0.05 2.41 0.23 0.06 30 0.87 0.49 0.10 2.91 0.43 0.09 35 1.05 0.50 0.08 3.32 0.43 0.07 Dynamic 20 25 2.74 0.35 0.06 2.08 0.34 0.06 30 2.83 0.37 0.03 3.16 0.37 0.03 35 2.93 0.46 0.03 3.81 0.39 0.03 30 25 - - - - - - 30 2.28 0.15 0.08 1.68 0.21 0.10 35 2.41 0.30 0.08 2.14 0.25 0.06 5.3.1.1 Percentage of crossings The percentage of crossings was examined using a three-way repeated-measures ANOVA with intrasubject variables vehicle initial distance (3) and vehicle speed (2), and the inter-subject variable experimental approach (2), as factors (see Figure 5.7). The experimental approach did not significantly influence the participants’ percentage of crossings F(1, 58) = 0.07, η2 = 0.01, p = 0.79. There was a main effect of vehicle speed, F(1, 58) = 116.21, η2 = 0.67, p < 0.01, with higher values for the 20 km/h condition compared to the 30 km/h condition. There was also a main effect of vehicle initial distance, F(2, 116) = 90.99, η2 = 0.61, p < 0.01, with Bonferroni post hoc tests showing that crossing percentages increased significantly with the initial distance (p < 0.01). A significant speed × initial distance interaction was also found, F(2, 116) = 49.74, η2 = 0.46, p < 0.01. Bonferroni post hoc tests were conducted comparing initial distances within each level of speed. At 20 km/h, there were significant differences between all levels of distance (25 m / 30 m: p < 0.01, 25 m / 35 m: p < 0.01, 30 m / 35 m: p = 0.01). At 30 km/h significant differences were only found between 25 and 35 m (p = 0.04). The experimental approach × initial speed and experimental
CHAPTER 5 109 approach × initial distance interactions were also significant, F(1, 58) = 4.79, η2 = 0.08, p = 0.03, and F(2, 116) = 9.21, η2 = 0.14, p < 0.01, respectively. The Bonferroni post hoc tests conducted to compare each level of speed within the experimental approaches revealed significant differences between the percentage of crossings regarding the initial speeds of 20 and 30 km/h in both experimental approaches, static (p < 0.01), and dynamic (p < 0.01). In turn, in the dynamic approach, there were significant differences between the percentage of crossings regarding all the initial distances (25 m / 30 m: p < 0.01, 25 m / 35 m: p < 0.01, 30 m / 35 m: p < 0.01). The same was verified in the static experiment (25 m / 30 m: p < 0.01, 25 m / 35 m: p < 0.01, 30 m / 35 m: p = 0.05). The experimental approach × speed × initial distance interaction had also a significant effect on percentage of crossings, F(2, 116) = 12.79, η2 = 0.18, p < 0.01. Bonferroni post hoc tests were conducted comparing experimental approaches within each level of speed with each level of initial distance. Only when the vehicle approached the crosswalk at 20 km/h and from 35 m, the percentage of crossings were significantly higher for the dynamic approach than for the static one (p < 0.01). In both experimental approaches, participants crossed more when vehicle speed was lower, and its initial position was farther from the crosswalk. At 20 km/h, the increase in distance resulted in greater growth of the crossings percentage, being even more evident in the participants’ responses that performed the dynamic experiment (see Figure 5.7). However, these percentages do not mean properly safe crossings. One advantage of the dynamic approach was to allow the exact determination of the occurrence of crashes. Considering the values presented in Table 5.4, it is possible to note that a considerable portion of the crossings made by participants resulted in a crash, particularly when the vehicle approached at 20 km/h from the shorter distance and 30 km/h from 30 m.
COMPARISON BETWEEN TWO TYPES OF EXPERIMENTAL APPROACH TO ASSESS THE PEDESTRIAN BEHAVIOUR IN VIRTUAL ENVIRONMENT 110 Figure 5.7 – Percentage of crossings and respective mean standard error as a function of experimental approach, per initial distance and initial speed, for Constant speed pattern. 5.3.1.2 Crossing start time Visual inspection of the residual plots showed deviations from homoscedasticity and skewness of crossing start time distribution, so the model considered to analyse this variable was fitted applying a logarithmic transformation to crossing start time, which corrected the deviations. In this way, an LMM of log(crossing start time) with random effects included for the participant and fixed effects defined for the initial speed , initial distance , and experimental approach was fitted. Satterthwaite’s tests showed significant effects of experimental approach, F(1, 53.08) = 46.40, p < 0.01, and initial speed, F(1, 790.60) = 3.77, p = 0.05. There was no effect of the initial distance, nor even of any interaction between the considered variables. In the static approach participants were quicker to start the crossing than in dynamic approach, b = - 0.90, t(58.79) = - 6.57, p < 0.01. When the vehicle approached the crosswalk at 30 km/h, participants started to cross sooner, b = - 0.24, t(789.50) = - 2.11, p = 0.04. Considering these results and Figure 5.8 that shows the crossing start time as a function of vehicle initial distance, initial speed, and experimental approach, it is noticeable that the participants’ decisions to cross tended to be faster at the highest speed and when they had not walked to perform the crossing task.
CHAPTER 5 111 Figure 5.8 – Crossing start time and respective mean standard error as a function of experimental approach, per initial distance and initial speed, for Constant speed pattern. 5.3.1.3 Time-to-passage The observed TTP appears to vary linearly with the initial TTP, particularly for the results of the static experiment, as already seen in the study presented in Chapter 4 (Figure 4.8), due to the small variations of the crossing start time. The crossing start time analysis done in the previous section showed that for those participants that did cross with the approaching vehicle at 30 km/h, there was an attempt to compensate for the shorter available time. However, this was not enough to compensate for the difference in TTP with the crossing TTP being almost linearly predicted by the Initial TTP. This is clearer for the results of the static experiment than for the dynamic one (Figure 5.9). Nevertheless, such as presented in Table 5.5, in the dynamic experiment, there were considerable percentages of crashes, particularly for the speed of 20 km/h with which the vehicle started its movement from 25 m far the crosswalk and for the speed of 30 km/h from the initial distance of 30 m. Disregarding the situations where a crash has occurred, it is possible to note, through the analysis of Figure 5.10, that, also in the dynamic approach, the crossing TTP can be almost linearly predicted by the Initial TTP. Furthermore, due to the later crossing start, the participants took riskier decisions in the dynamic approach.
COMPARISON BETWEEN TWO TYPES OF EXPERIMENTAL APPROACH TO ASSESS THE PEDESTRIAN BEHAVIOUR IN VIRTUAL ENVIRONMENT 112 Figure 5.9 – Mean values of TTP at the crossing instant as a function of the TTP at the start of the trial, by experimental approach, for Constant speed pattern. Figure 5.10 – Mean values of TTP at the crossing instant as a function of the TTP at the start of the trial, disregarding crashes, by experimental approach, for Constant speed pattern. 5.3.2 Different speed patterns Figure 5.11(a), Figure 5.11(b), Figure 5.12(a) and Figure 5.12(b) show the crossing data for each participant and condition (location of the circle along the axis represents the mean crossing start time, the size of the circle depicts the number of crossings, and the line shows the standard error of the mean) and the speed profile of the vehicle. It is possible to note that the number of crossings increases substantially for the Slow Down and Stop patterns in both experimental approaches, being this particularly
CHAPTER 5 113 clearer in the dynamic approach, in which very few participants had crossed when the vehicle approached them at a constant speed of 30 km/h. Table 5.6 and Table 5.7 show the summary of descriptive statistics for the condition in terms of percentage of crossings, percentage of crashes, crossing start time, and TTP regarding the different speed patterns. (a) (b) Figure 5.11 – Participants’ responses along time of stimulus’ presentation, regarding the initial speed of 20 km/h, per speed pattern and experimental approach: (a) dynamic; and(b) static.
COMPARISON BETWEEN TWO TYPES OF EXPERIMENTAL APPROACH TO ASSESS THE PEDESTRIAN BEHAVIOUR IN VIRTUAL ENVIRONMENT 120 Figure 5.15 – TTP and respective mean standard error as a function of experimental approach, per initial speed and speed pattern. 5.3.3 Comparison between virtual and real environments An LMM of TTP with subject ID as a random effect and the speed pattern and the data gathering approach as fixed factors were fitted to compare the results obtained with each experimental approach’s performance and video recordings. Satterthwaite’s tests just showed significant speed pattern effects, F(2, 994.40) = 5.15, p < 0.01. There was no significant effect of the data gathering approach, F(2, 106.69) = 0.82, p = 0.44, nor of the data gathering approach × speed pattern interaction, F(4, 1432.49) = 1.90, p = 0.11. The model was refitted, removing the experimental approach factor, and contrasts were used to compare the different speed patterns. A first contrast compared the Constant and Slow Down speed patterns and, such as in the previous section, non-significant differences were found between them. A second contrast compared the Constant and Stop patterns and it showed a significant difference between them, b = 1.35, t(2769.94) = 5.47, p < 0.01. A third contrast has compared the Slow Down and Stop patterns and also revealed a significant difference between them, b = 1.74, t(2747.00) = 7.72, p < 0.01. These results meet those ones presented in section 5.3.2.3, demonstrating that the TTP was in general longer, not only in virtual environment experiments but also in real crossing situations, for the Stop pattern. Furthermore, the statistical analysis also revealed that the different data collection methods did not induce substantial differences in how pedestrians estimate the time they need to cross the road safely. Figure 5.16 shows the referred similarities between the three methods. Although the results of the LMM
CHAPTER 5 121 do not show it, a slight difference can be noticed when comparing the static approach with the other two, particularly in the TTP observed in the Constant Speed pattern condition. Figure 5.16 – Comparison of the TTP obtained in real and virtual environment through the execution of each experimental approach, per speed pattern. 5.4. Discussion As referred in the introductory section of this chapter, according to Feldstein et al. (2016), the capacity of making participants feel they are actually present in the virtual environment determines the realism of the environment and the usability of the simulator. This will depend on the quality of the graphical representation, sound, and interaction possibilities. This study compares the results of the experimental approach used in Chapter 4, where participants had their movements entirely restricted, with the ones obtained through an experimental approach that allowed them to walk along a crosswalk, aiming to assess the impact of the interaction between the participant and virtual environment on crossing decision-making. Regarding of the crossing decisions observed when the vehicle approached the participants at a constant speed, and congruently with the results obtained in Chapter 4, the vehicle speed and initial distance were the most determining variables for the participants’ crossing decisions. The percentage of crossings only differed between the two experimental approaches when the vehicle approached at 20 km/h from the
COMPARISON BETWEEN TWO TYPES OF EXPERIMENTAL APPROACH TO ASSESS THE PEDESTRIAN BEHAVIOUR IN VIRTUAL ENVIRONMENT 122 initial distance of 35 m, i.e. , in the most favourable condition to cross and in which the percentage of crossings was higher. Such as in the static experimental approach, in the dynamic method, crossing percentages increased with vehicle initial distance and decreased with its speed. These results confirm again the important role of vehicle speed in crossing decisions when balanced with the distance weight, contrasting with Feldstein and Dyszak (2020) results. Crossing start times were not affected by the initial distance. However, the vehicle speed and the experimental approach had a significant impact on participants’ time to decide to cross. In both experimental approaches, higher speeds have prompted faster decisions. As for the experimental approach’s effect, it can be easily explained by the distance of 3 m that participants had to walk after the beginning of the stimuli presentation and before reaching the curb in the dynamic experiment. Except for the condition characterized by the speed of 30 km/h and the initial distance of 25 m, where none crossing in the dynamic approach, the crossing start times assume the same trend in both experimental methods. In the dynamic approach, the crossing start times were, on average, 1.58 s higher than in the static experiment. For this reason, and for both experimental approaches, it is possible to affirm that shorter stimuli led to less time to take the decision, agreeing with Lobjois and Cavallo (2009). Again, the higher speeds and close distances have prompted a similar sense of urgency, accelerating the decision processes. This had a repercussion on the accuracy of judgements. In the dynamic experiment, where it was possible to determine the number of crashes, considering the values presented in Table 5.4, they expressively occurred when vehicle speed was the highest and when the vehicle approached the crosswalk at 20 km/h from the shortest distance. Thus, it was possible to note, with both experimental approaches, that the TTP and the crossing start time were directly related when the vehicle approached at a constant speed. Regarding the analysis of the stimuli considering the three different types of speed pattern, the results showed that the initial speed of the vehicle had a significant influence both on the percentage of crossings, with the participants crossing more often when the initial speed of the vehicle was 20 km/h, and on the crossing start time. This is valid particularly for the dynamic experiment, since in Chapter 4, through the execution of the static experiment, it was verified that crossing start time (response time) was significantly longer when the vehicle approached the crosswalk at an initial speed of 30 km/h.
CHAPTER 5 123 In general, participants crossed more when the vehicle varied its speed (Slow Down and Stop patterns). In these patterns, most of the static experiment participants and all the participants of the dynamic experiment waited for a speed considerably lower than the initial speed to cross the virtual road. The crossing start time was longer in the dynamic compared to the static experiment. However, as in the constant speed pattern, the delay in crossing decision-making verified in the dynamic experiment is defined by the distance the participants had to walk before arriving at the crosswalk and not by a better ponderation made before the crossing, as indicated by similarity between the percentage of crossings for both experimental approaches. The TTP analysis showed that, for both experimental approaches only the condition characterized by the Stop Pattern with an initial speed of 30 km/h was significantly different from the others. For this condition, the participants crossed mostly when the vehicle speed was almost 0 km/h, making the TTP higher than in the other conditions. The TTP values for the different speed patterns analysis confirm the existing similarity between both experimental approaches. Considering the general comparison between the results of the static and dynamic experiments with those obtained through the analysis of the two videos recorded in the real environment, it was possible to note that there was no significant difference between them. Thus, and considering all the advantages of carrying out experiments in virtual environments, (Deb et al. , 2017), this study confirms that the use of a simulator, regardless of the practical experimental approach, is a sustainable option to take in the pedestrian safety research area. On the other hand, the choice of the approach to be implemented in each study must depend on the desired amount of information to get. With the possibility of extracting participants’ trajectories and speeds, the dynamic approach allows calculating the most various surrogate safety indicators ( e.g. , PET, TTCmin, TA, etc. (Johnsson et al. , 2018a)), contrary to the static approach. Besides, it allows to perform the crossing task in a similar way to that occurring in the real world. However, an experience where the participants walk along a crosswalk is technically more demanding than the one where they click on a button when they decide to cross, because in terms of development effort, it is more complex to model and implement the sound and the visual scenarios. All these characteristics must be well pondered in the design phase of each study. The static experience can be more effective when applied in studies where
COMPARISON BETWEEN TWO TYPES OF EXPERIMENTAL APPROACH TO ASSESS THE PEDESTRIAN BEHAVIOUR IN VIRTUAL ENVIRONMENT 124 a great detail of information is unnecessary, while the dynamic experience can be useful for more indepth studies. 5.5. Conclusions Such as the sound, the possibility of letting participants move freely turns the simulator more realistic and immersive, allowing them to have a more complete interaction with the virtual world. In this way, this work aimed to assess the implementation of a more realistic approach for studying pedestrian crossing behaviour, comparing two different experimental approaches: (i) the static approach used in the studies presented in Chapter 3 and Chapter 4, in which the participants were required to decide when they would cross the road by clicking on a button, standing in the same position during all the experiment; (ii) the dynamic approach, in which the participants were instructed to cross the virtual road, walking along a semi-virtual crosswalk. The overall analysis reveals that the experimental approach was not a determinant factor on the participants’ crossing decision task. As in Chapter 3 and Chapter 4, the obtained results support the general conclusion that the vehicle speed and initial distance, as well speed profile, were the variables used by participants to make their decision. The static approach has the advantage of turning the experimental task simpler and less time consuming, with instructions easily assimilated and performed by the participants. The dynamic is more timeconsuming due to the circuit that pedestrians must walk to answer each of the stimuli presented. However, it is more naturalistic than the static experiment. It allows gathering a greater quantity of information, such as participants’ speed and position and the determination of crash occurrence. Regardless of these main characteristics, both experimental approaches revealed to be valid for studying the pedestrians’ crossing decision-making. The use of each of the approaches in future studies must be considered depending on the desired type of information and the detail intended.
125 6. ANALYSIS OF PEDESTRIAN BEHAVIOUR IN VIRTUAL ENVIRONMENT 6.1. Introduction Several studies addressing the pedestrians’ behaviour when crossing the road have been carried out to identify factors that can affect and influence it, as showed in Chapter 1 and Chapter 2. Identifying factors that lead to risky behaviour is a way to facilitate and increase the effectiveness of public policies aimed to solve the big problem that is the high accident rate and, consequently, pedestrian mortality on the roads. As simulators become more available to researchers, some studies have been carried out using virtual environments as a tool to collect data, which enables a finer analysis of participants’ behaviour and decision-making ( e.g. (Cavallo et al. , 2019; Charron et al. , 2012; Deb et al. , 2018b; Feldstein et al. , 2016; Meir et al. , 2015; Morrongiello et al. , 2015; Schwebel et al. , 2008; Simpson et al. , 2003; Sween et al. , 2017; Thomson et al. , 2005; Zito et al. , 2015)). This has been the main method of study reported in Chapters 3, 4, and 5 of this document. Simulators allow a greater control over the variables considered in a study, in addition to allowing crossings without jeopardizing the safety and physical integrity of the participants (Cavallo et al. , 2019; Charron et al. , 2012; Deb et al. , 2017; Dommes and Cavallo, 2011; Schwebel et al. , 2012; Simpson et al. , 2003). This makes them a very interesting and appealing tool to use in studies concerning the identification of impacting factors related to road infrastructure and the built environment in pedestrian crossing decision-making. Factors identified in the literature, such as the road width, the number of lanes, the width and quality of the sidewalks, the parking spaces, among other physical characteristics of the simulated streets, can be easily manipulated from the point of view of their dimensions. Different variations can be integrated into the virtual scenarios to be presented in each experiment. This part of the study, which comes as a sequence of all the work carried out and presented in the previous chapters, takes advantage of the simulator developed throughout the ANPEB project, within which this doctoral project was developed, to complement the results obtained in the analysis of
ANALYSIS OF PEDESTRIAN BEHAVIOUR IN VIRTUAL ENVIRONMENT 126 pedestrian behaviour in the real environments (Chapter 2), through the application of the experimental protocol developed and tested with the execution of the experiments described on Chapter 5. However, in this case, four new virtual scenarios were added to the two implemented in the study presented on the previous chapter, aiming to assess the influence of a set of factors related to the characteristics of the road and pedestrian infrastructure, without disregarding the participants’ age and sex, as well as the speed of approach of the vehicle, in the participants’ crossing decision-making, through the construction of two models: one for TTP and other fot TTCmin. 6.2. Materials and Methods 6.2.1 Participants A sample of 45 adults were recruited from the University of Minho community, in Portugal. One third of them were part of the sample of participants that performed the dynamic experiment presented in Chapter 5. The participants were divided into three distinct groups. To each group was associated a different pair of virtual scenarios where they performed the experiment. The details about the demographic characteristics of the participants are presented in Table 6.1. Table 6.1 – Participants’ demographic characteristics. Group I II III Total Age 23 - 39 years 21 - 38 years 21 - 39 years 21 - 39 years (m = 30.00; (m = 29.00; (m = 27.40; (m = 28.80; sd = 4.95) sd = 5.45) sd = 6.08) sd = 5.62) Sex 53 % Male; 53 % Male; 47 % Male; 51 % Male; 47 % Female 47 % Female 53 % Female 49 % Female Also, in this study, before the experiment, all participants answered a questionnaire regarding their hearing, visual, and mobility conditions. None of them reported any impairing condition. All participants gave their written informed consent. The experiments were conducted in accordance with the principles stated in the 1964 Declaration of Helsinki.
CHAPTER 6 127 6.2.2 Virtual environment Three combinations of 15 participants and two scenarios were used in this study. By scenarios one means a particular virtual street (replicated from one of the real streets analysed in chapter 2). Each individual combination was presented to one of the three groups: - Group I: the scenarios of the 25A and TP streets used in the studies in the virtual environments previously presented (Chapter 3, Chapter 0 and Chapter 5) (see Figure 6.1(a) and Figure 6.1(b)); - Group II: the scenarios of the BMJ and SG streets (see Figure 6.1(c) and Figure 6.1(d)); - Group III: the scenarios of CL and AGC streets (see Figure 6.1(e) and Figure 6.1(f)). The scenarios’ modelling was done considering the real characteristics and dimensions of each street (Table 6.2). Table 6.2 – Main characteristics of each one of the six scenarios. Street 25A TP BMJ SG CL AGC Length of the crosswalk (m) 7.85 7.13 9.96 12.5 8.97 7.03 Average width of the lanes (m) 2.87 3.57 3.26 3.00 3.32 2.56 Width of the street to park (m) 4.23 10.12 1.92 5.17 2.34 7.15 Average width of the sidewalks (m) 4.17 4.11 1.52 2.96 1.32 1.95 Width of the crosswalk (m) 3.40 5.12 3.04 5.35 3.00 3.54 Number of lanes 2 2 2 4 2 2 The textures used to model the scenarios were obtained through photos and edited using Adobe® Photoshop CC 2015 before being included in the 3D model. The physical dimensions of the virtual elements were equalized to the real-world measurements. The method to model a vehicle’s movement was the same used in the studies in the virtual environments presented in the two previous chapters. Once again, it is important to mention that speeds above 30 km were not considered since this study aimed to evaluate the pedestrians’ crossing decision-making, considering the approaching of a vehicle at short distances from the crosswalk in different scenarios. As
ANALYSIS OF PEDESTRIAN BEHAVIOUR IN VIRTUAL ENVIRONMENT 128 previously referred, it was assumed that higher speeds would lead to the absence of useful data because participants would rarely cross. (a) (b) (c) (d) (e) (f) Figure 6.1 – The six scenarios considered in this study: (a) 25A; (b) TP; (c) BMJ; (d) SG; (e) CL; (f) AGC.
CHAPTER 6 129 6.2.3 Stimuli The same visual and auditory stimuli were presented to the three groups of participants. Only the visual scenarios changed. The ten conditions shown in Table 4.2 were repeated five times for every participant. Throughout the experiment, 100 stimuli were presented in a random order (10 movement conditions × 5 repetitions × 2 streets) to each participant. The virtual model of the approaching vehicle used in the experiment was the same used in the experimental task performed in the study of Chapter 5 (see Figure 5.1). The auditory component of the stimuli was composed of the same auralized CPX sounds used in the experiments presented in Chapter 5 (see section 5.2.2.). 6.2.4 Instruments The experiment was conducted in the same room and using the same CAVE type system used in the studies described in the three previous chapters (see section 3.2.4). 6.2.5 Experimental procedure The experimental procedure adopted in this study concerns the same used in the dynamic approach described in section 5.2.5 of Chapter 5. Equipped with the headset and the 3D glasses, participants were instructed to walk along the predefined circuit shown in Figure 5.3 and cross the virtual crosswalk when they felt safe to do so. As in the other experiments, participants completed an experimental session made of two main blocks, each one using one of the two respective scenarios, preceded by a training block composed of 4 stimuli. There was a gap of 5 minutes or more between the two main blocks to rest, depending on the participant’s wishes. 6.2.6 Analysis The street scenario’s influence on the participants’ crossing decision-making was analysed in terms of the percentage of crossings, percentage of crashes, crossing start time, TTP, and TTCmin. As in Chapter 5,