Many Ways Lead to the Goal—Possibilities of Autonomous and Infrastructure-Based Indoor Positioning

electronics Article Many Ways Lead to the Goal—Possibilities of Autonomous and Infrastructure-Based Indoor Positioning Hossein Shoushtari 1,*, Thomas Willemsen 2and Harald Sternberg 1   Citation: Shoushtari, H.; Willemsen, T.; Sternberg, H. Many Ways Lead to the Goal— Possibilities of Autonomous and Infrastructure-Based Indoor Positioning. Electronics 2021,10, 397. https://doi.org/10.3390/ electronics10040397 Academic Editor: Raed A. Abd-Alhameed Received: 14 December 2020 Accepted: 2 February 2021 Published: 5 February 2021 Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations. Copyright: © 2021 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/). 1HafenCity University, 20457 Hamburg, Germany; harald.sternber[email protected] 2Hochschule Neubrandenburg, 17033 Neubrandenburg, Germany; [email protected] *Correspondence: [email protected] Abstract: There are many ways to navigate in Global Navigation Satellite System-(GNSS) shaded areas. Reliable indoor pedestrian navigation has been a central aim of technology researchers in recent years; however, there still exist open challenges requiring re-examination and evaluation. In this paper, a novel dataset is used to evaluate common approaches for autonomous and infrastructurebased positioning methods. The autonomous variant is the most cost-effective realization; however, realizations using the real test data demonstrate that the use of only autonomous solutions cannot always provide a robust solution. Therefore, correction through the use of infrastructure-based position estimation based on smartphone technology is discussed. This approach invokes the minimum cost when using existing infrastructure, whereby Pedestrian Dead Reckoning (PDR) forms the basis of the autonomous position estimation. Realizations with Particle Filters (PF) and a topological approach are presented and discussed. Floor plans and routing graphs are used, in this case, to support PDR positioning. The results show that the positioning model loses stability after a given period of time. Fifth Generation (5G) mobile networks can enable this feature, as well as a massive number of use-cases, which would benefit from user position data. Therefore, a fusion concept of PDR and 5G is presented, the benefit of which is demonstrated using the simulated data. Subsequently, the first implementation of PDR with 5G positioning using PF is carried out. Keywords: indoor navigation; autonomous; infrastructure; particle filter; pedestrian dead reckoning; fusion 5G; inertial sensors 1. Introduction The development of reliable indoor positioning based on Micro-Electro-Mechanical Sensors (MEMS) has been a central aim of technology researchers in recent years. Not only are such sensors cost and energy effective, but they also correspond to Mark Weiser’s vision [ 1 ]; in that they work for everyone, anywhere. Thus, the objective is to develop applications that are as embedded as possible. Indeed, every smartphone— carried around by almost everyone today—can be counted as a unique source of such an application [2]. Another possibility would be 5G New Radio (NR) networks, which are expected to enable highly accurate positioning as an ideal support for sensor-based autonomous approaches. They will be available to the vast majority of the population, as is planned for 2027 in the U.K., for example [ 3 ]. The high dependency between smartphones and 5G signals seems a perfect match. Moreover, wireless radio-based communications have the intrinsic problem of signal unreliability, due to multipath effects and No Line-of-Sight signal conditions, particularly in indoor environments [ 4 ], which can be improved by using the abovementioned sensors. Inertial measurement units (IMU) are such sensors, which are not highly affected by environmental conditions; thus, many modern positioning techniques rely on their use as key sensors. The aim is to target hundreds of location-based services while keeping humans as a central part of the system, using sensing technologies such as wearable sensors, which Electronics 2021,10, 397. https://doi.org/10.3390/electronics10040397 https://www.mdpi.com/journal/electronics Electronics 2021,10, 397 2 of 17 have been trending in application [ 5 ]. Tracking the human pose by using virtual reality headsets, for example, is another of many examples of such applications. Many smart objects, such as watches and shoes, can also directly benefit pedestrian tracking techniques, especially in GNSS-shaded areas where there is no guiding reference. To improve localization accuracy, prior works have generalized and fused other sources which utilize descriptive observations. The observations from different sensors, such as vision features [ 6 , 7 ], point clouds [ 8 , 9 ] from cameras, and Laser Identification Detection and Ranging (LIDAR) sensors, can be combined with odometry information to correct for drift and to provide efficient localization and tracking at the same time [ 10 – 12 ]. However, such vision-based auxiliary features are not always available. For instance, visual features fail in dark environments and LIDAR features fail in heavily reflective environments. In this paper, we revisit the fundamental question: If only smartphone data (including the 5G NR network) and an environment map is provided, is it possible to localize a pedestrian for a long period of time? To achieve this, map support is not enough. Different possibilities are discussed, and it is shown that the special performance of the future mobile communication standard 5G NR plays a central role as an enabling technology. Therefore, 5G positioning, which is an infrastructure-based solution that provides absolute position information, is considered. The 5G technology is used in our final approaches, thus showing that 5G can potentially serve as a base requirement. Dense NR networks are envisioned to range from a few meters up to tens of meters; for example, assuming several access nodes per room in indoor environments [ 13 ]. Thus, 5G NR, as an infrastructurebased positioning solution, can be made available both indoors and outdoors and is an ideal way to navigate in cities (e.g., between buildings, subways, and train stations). The remainder of the paper is structured as follows: the related work and challenges are described in the next section, section three presents the study dataset, in section four, two main positioning techniques which can work independently—that is, only smartphone sensors and a simulation for Infrastructure-based implementation (e.g., 5G)—are discussed; section five presents the autonomous position estimation method, where particle filtering map-matching is discussed; in section six, we present an initial fusion concept for the autonomous approach and 5G—which can be realized with and without taking mapmatching into consideration; and finally, section seven provides our conclusions and directions for future works. 2. Related Work and Challenges Over time, indoor positioning methods have evolved to the current stage, which now show an accuracy performance up to the level of one meter. However, the positioning techniques are still mostly limited to sensor error accumulation, device placements or finite labeled data. A wide range of different sensor technologies have been used in order to overcome these limits during the last few years, especially through smartphones, which have almost all the information one needs in one place. Within the available sensors in smartphones, inertial measurement units (IMUs) are in demand. This is largely because IMUs contain sensors (such as accelerometers and gyroscopes) and they provide information about the orientation and position of any object that they are attached to. However, different applications and environmental conditions influence the sensor selection and fusion type. Figure 1shows a historical timeline of the smartphone-based positioning techniques, beginning with accelerometers. The positioning must be accurate and reliable, even when the use of the device is varied in terms of placement (e.g., handheld or in a packet) or orientation (e.g., portrait or landscape mode). Authors in [ 14 ] have tried to combine different physics-based methods to overcome placement challenges, such as the Inertial Navigation System (INS), Pedestrian Dead Reckoning (PDR), or different filtering approaches such as the Kalman Filter (KF) for orientation tracking and the Particle Filter (PF) for map-matching. However, they could not argue that all can be done in real time applications. Therefore, open challenges and Electronics 2021,10, 397 3 of 17 questions can be defined clearly based on the timeline review and by aiming for a robust and real time positioning solution. Electronics 2021, 10, x FOR PEER REVIEW 3 of 17 (KF) for orientation tracking and the Particle Filter (PF) for map-matching. However, they could not argue that all can be done in real time applications. Therefore, open challenges and questions can be defined clearly based on the timeline review and by aiming for a robust and real time positioning solution. Figure 1. The timeline of smartphone-based positioning techniques obtained from extensive literature review [15–29]. A research challenge regarding pose tracking is how to determine the initial pose of the device. This is a very important question—if the initial position calculation is possible, then it should be also possible to estimate the new position for the next points. The pedestrian navigation should be comparable in use to GNSS-based navigation, in a way that the approach works with every smartphone in anytime. However, in order to provide highquality positioning, absolute methods using Wireless Local Area Network (WLAN), Bluetooth, or Ultra-Wide Band (UWB) are among those individual solutions which often require big infrastructural changes, for example the additional mount of access points in fingerprint approaches [18,20,30]. GPS initialization [19] has been accepted as a standard initialization method; however, it obviously imposes a strong limit on the starting point. Manual user input is the only remaining way to start a navigation journey, perhaps by using a QR code as another specialized solution [23]. The initial orientation has fewer problems, as pitch and roll are calculable based on the gravity estimation from accelerometer measures [30]. Calibrated geomagnetic sensor measurement following a local magnetic map can also help in this process, but this is highly reliant on a labeled dataset [22]. The next challenge regards pedestrian positioning. Many studies on orientation tracking in the context of indoor positioning have attempted to estimate the heading of the device, rather than the heading of the pedestrian. They, therefore, generally assume that the position where the mobile device is attached is fixed and known, or they neglect the misalignment between smartphone heading and the user’s walking direction, leaving it as an open challenge [31]. However, there have been some attempts to calculate the heading independent of orientation, either following the very first attempts based on acceleration values and PCA calculation [32], or similar approaches such as frequency domain analysis; however, such methods must be improved to reach the desired accuracy [33]. The last two challenges become even more problematic when the placement of the device varies over time as a user performs different tasks, such as making a call, carrying the device in a bag, or walking on an escalator. There are also limitless human behaviors, such as side walking, back walking, strange or jump stepping, taking stairs two at a time, and many other possibilities. Researchers have recently tried to acquired sensor data across a limited number of human subjects and smartphone placements. In this manner, the first supervised training dataset for inertial navigation was introduced in 2017, using one Support Vector Machine (SVM) and eight Support Vector Regression (SVR) models [27]. Realistic and open benchmarking datasets for pedestrian indoor positioning using Figure 1. The timeline of smartphone-based positioning techniques obtained from extensive literature review [15–29]. A research challenge regarding pose tracking is how to determine the initial pose of the device. This is a very important question—if the initial position calculation is possible, then it should be also possible to estimate the new position for the next points. The pedestrian navigation should be comparable in use to GNSS-based navigation, in a way that the approach works with every smartphone in anytime. However, in order to provide high-quality positioning, absolute methods using Wireless Local Area Network (WLAN), Bluetooth, or Ultra-Wide Band (UWB) are among those individual solutions which often require big infrastructural changes, for example the additional mount of access points in fingerprint approaches [ 18 , 20 , 30 ]. GPS initialization [ 19 ] has been accepted as a standard initialization method; however, it obviously imposes a strong limit on the starting point. Manual user input is the only remaining way to start a navigation journey, perhaps by using a QR code as another specialized solution [ 23 ]. The initial orientation has fewer problems, as pitch and roll are calculable based on the gravity estimation from accelerometer measures [ 30 ]. Calibrated geomagnetic sensor measurement following a local magnetic map can also help in this process, but this is highly reliant on a labeled dataset [22]. The next challenge regards pedestrian positioning. Many studies on orientation tracking in the context of indoor positioning have attempted to estimate the heading of the device, rather than the heading of the pedestrian. They, therefore, generally assume that the position where the mobile device is attached is fixed and known, or they neglect the misalignment between smartphone heading and the user’s walking direction, leaving it as an open challenge [ 31 ]. However, there have been some attempts to calculate the heading independent of orientation, either following the very first attempts based on acceleration values and PCA calculation [ 32 ], or similar approaches such as frequency domain analysis; however, such methods must be improved to reach the desired accuracy [33]. The last two challenges become even more problematic when the placement of the device varies over time as a user performs different tasks, such as making a call, carrying the device in a bag, or walking on an escalator. There are also limitless human behaviors, such as side walking, back walking, strange or jump stepping, taking stairs two at a time, and many other possibilities. Researchers have recently tried to acquired sensor data across a limited number of human subjects and smartphone placements. In this manner, the first supervised training dataset for inertial navigation was introduced in 2017, using one Support Vector Machine (SVM) and eight Support Vector Regression (SVR) models [ 27 ]. Realistic and open benchmarking datasets for pedestrian indoor positioning using modernday smartphones together with ground-truth has become a recent research activity [ 28 , 29 ], employing deep learning-based approaches [ 2 , 26 ]. It has been found that relying on limited labeled data sources can lead to an end result. These authors did not discuss how realistic the dataset could be, due to the experimental situation; however, the question is: How can Electronics 2021,10, 397 4 of 17 a dataset be created, which well-represents a user acting freely, normally, and away from the experimental situation? In order to establish an unlimited positioning approach capable of operations everywhere and at any time, as well as to find a solution for the abovementioned challenges, the potential of 5G NR technology can be considered as a key requirement. Owing to the previous popularity of smartphones for positioning estimation, 5G-based methods could be a perfect match for them. Some researchers have reported reaching an accuracy of one meter or better for 70% of the usage time different positioning measurements, such as Time of Arrival (ToA) and Direction of Arrival (DoA), in combination with Kalman Filtering [ 13 ]. 3. The Study Dataset This paper contributes a dataset of smartphone sensor measurements and the results of a 5G-based simulation. It also includes the required map information and interpreted Python scripts. We used a handheld Samsung S10 5G smartphone (Samsung, Suwon, Gyeonggi-do, Korea) [ 34 ], to record acceleration, magnetic field, barometer and orientation sensor values, as well as angular velocities by the gyroscope via an application named “sensor log” [ 35 ]. The simulation result, discussed in Section 4.2.2, is presented along with the building plan (in WKT format). The 5G simulator—in combination with manual video controls—was used to estimate a good ground truth trajectory. The reader script read the sensor log file and extracted the data with desired frequencies. We solely used IMU data at 100 Hz for all of the following approaches. For the autonomous approaches, the initial point was manually set to be 566578.7 and 5932830.1 meters for the initial position and 200 degrees for initial heading. The dataset is available online at the following link (https://github.com/Hossein-Shoushtari/ElectronicsData.git) (accessed on 15 December 2020). It should be noted that the height component was omitted in the implementations. 4. Selection of Positioning Techniques In the following section, two general techniques which provide position information sources are discussed. The first one only relies on sensor information, while the other just uses infrastructure measurements. 4.1. Pedestrian Dead Reckoning The accelerometer–gyroscope combination is of particular interest for pedestrian dead reckoning. These MEMS inertial sensors are often combined as three-axis sensors in smartphones, such as the MPU-9250 (TDK InvenSense, Tokyo, Japan) [ 36 ]. When the smartphone is in resting position, the acceleration sensor registers acceleration due to gravity on its three axes (i.e., 9.81 m/s 2 ). This allows for tilt calculation of rotation in the smartphone–sensor coordinate frames. Step detection is also based on the registered acceleration. The three-axis gyroscope registers angular velocities. With appropriate time measurements, the relative angles of rotation in space can be determined by integration. The inertial sensors are subject to colored noise, which is even more apparent in MEMS. This can be seen in the angle of rotation, due to drift. The temperature dependence also has a great influence. These effects produce an increase of positioning error along with time of use in PDR. Therefore, the fusion of gyroscope data (drift in angle) and acceleration data (high noise in angle) helps to reduce uncertainties for two or three orientation parameters. For instance, Gyroscope and Acceleration sensor fusion has been made possible by means of the well-known Madgwick algorithm [37]. The basis of PDR is the combination of a pedometer, known or estimated stride length, and orientation. There are numerous implementation examples for the realization of the PDR, in which the positioning—both in terms of step detection and heading estimation— must be accurate and reliable. Once a step is detected, the system needs to estimate the stride length. The stride length of an individual can vary significantly over time, due to speed, terrain, and other environmental constraints [ 25 ]. The same problem applies to Electronics 2021,10, 397 5 of 17 the exact heading direction for each step, despite the fact that the phone can be located anywhere on the body. Using the benchmarking methods from [ 21 ] and [ 25 ], two different PDR realizations were reproduced (see Figures 1and 2). These figures show the results of a real trajectory dataset in a building of HafenCity University. Electronics 2021, 10, x FOR PEER REVIEW 5 of 17 The basis of PDR is the combination of a pedometer, known or estimated stride length, and orientation. There are numerous implementation examples for the realization of the PDR, in which the positioning—both in terms of step detection and heading estimation—must be accurate and reliable. Once a step is detected, the system needs to estimate the stride length. The stride length of an individual can vary significantly over time, due to speed, terrain, and other environmental constraints [25]. The same problem applies to the exact heading direction for each step, despite the fact that the phone can be located anywhere on the body. Using the benchmarking methods from [21] and [25], two different PDR realizations were reproduced (see Figures 1 and 2). These figures show the results of a real trajectory dataset in a building of HafenCity University. (a) (b) Figure 2. Realization of PDR elements reproducing available benchmarks, listed as: (a) Step heading calculation using Madgwick fusion of IMU and leveled Gyroscope considering ZUPT; and (b) K-based stride length estimation. Figure 2 shows the use of Madgwick fusion of IMU data to azimuth, implemented using a Euler free quaternion-based method and the integration of gyroscope data through ZeroVelocityUPdaTe (ZUPT: Offset of rotation rate to Zero at standstill) at the beginning. Based on both of these examples, a difference of more than 20° is visible; however, these results are based on the same data. They make it clear that the selection of orientation estimation method has a significant influence on position estimation. However, one disadvantage of Euler based systems is the singularity limitation. Stride length estimation based on accelerometer data works dynamically and there is no need for manual input, in comparison with the use of a fixed stride length. Based on the experiments, the k-based method generally works better for different users without changing the set of parameters. However, a variation of 20 cm in stride length between two individual’s strides in a building is unrealistic. Moreover, the correction of a fixed stride length can be done more easily and quickly. The reliable determination of stride length requires a realistic dataset and learning approaches. Therefore, it can be still be seen as an unsolved challenge. In [21], the authors concluded that a fixed pre-estimated stride length causes the same deviations for an individual as the k-based methods. The next question regards the pedometer. The authors in [38] used a query of two successive conditions for the pedometer. For this,  with 9.81 m/s 2 had to be subtracted from the levelled accelerometer data and two thresholds for the acceleration value and one for the time were used. Other similar threshold-based pedometers have been presented in the literature (see, e.g., [25]). However, the robustness of the step detector methods remains questionable. The aforementioned reference assumed that people generate a periodic acceleration signal only when walking and could not distinguish between a real step and actions that generate very similar acceleration signal patterns, such as shaking the smartphone [30]. This may pose considerable problems, considering human daily behaviors such as walking on stairs, standing, sitting, sending texts, making calls, and typing. We seek to resolve this challenge by considering both horizontal and vertical displacement in the fusion concept introduced in Section 5. Figure 2. Realization of PDR elements reproducing available benchmarks, listed as: ( a ) Step heading calculation using Madgwick fusion of IMU and leveled Gyroscope considering ZUPT; and ( b ) K-based stride length estimation. Figure 2shows the use of Madgwick fusion of IMU data to azimuth, implemented using a Euler free quaternion-based method and the integration of gyroscope data through ZeroVelocityUPdaTe (ZUPT: Offset of rotation rate to Zero at standstill) at the beginning. Based on both of these examples, a difference of more than 20 ◦ is visible; however, these results are based on the same data. They make it clear that the selection of orientation estimation method has a significant influence on position estimation. However, one disadvantage of Euler based systems is the singularity limitation. Stride length estimation based on accelerometer data works dynamically and there is no need for manual input, in comparison with the use of a fixed stride length. Based on the experiments, the k-based method generally works better for different users without changing the set of parameters. However, a variation of 20 cm in stride length between two individual’s strides in a building is unrealistic. Moreover, the correction of a fixed stride length can be done more easily and quickly. The reliable determination of stride length requires a realistic dataset and learning approaches. Therefore, it can be still be seen as an unsolved challenge. In [ 21 ], the authors concluded that a fixed pre-estimated stride length causes the same deviations for an individual as the k-based methods. The next question regards the pedometer. The authors in [ 38 ] used a query of two successive conditions for the pedometer. For this, g with 9.81 m/s 2 had to be subtracted from the levelled accelerometer data and two thresholds for the acceleration value and one for the time were used. Other similar threshold-based pedometers have been presented in the literature (see, e.g., [ 25 ]). However, the robustness of the step detector methods remains questionable. The aforementioned reference assumed that people generate a periodic acceleration signal only when walking and could not distinguish between a real step and actions that generate very similar acceleration signal patterns, such as shaking the smartphone [ 30 ]. This may pose considerable problems, considering human daily behaviors such as walking on stairs, standing, sitting, sending texts, making calls, and typing. We seek to resolve this challenge by considering both horizontal and vertical displacement in the fusion concept introduced in Section 5. Finally, PDR is then performed for each detected step. In Equation (1), the coordinates of the current position Xi are calculated, depending on the previous position Xi−1 , summing with a multiplication of the scalar translation of T (stride length), and heading matrix R. Xi=Xi−1+T×R (1) The two PDR realizations led to different trajectories (see Figure 3). The trajectories were both subject to significant orientation drift and scale errors caused by incorrect stride Electronics 2021,10, 397 6 of 17 length estimation. In order to evaluate and track the approaches quality on-board, a cumulative distribution function (CDF) plot was placed next to each of the estimated trajectories. The CDFs were calculated by fitting a normal distribution to distances between the ground truth step coordinates and the estimated one, i.e., errors. In other words, the error is defined as the nominal value minus the estimated value. If there was a contrast between the pedometer and the actual number of the steps, a simple interpolation reduced or added a few step positions. The evaluation of the PDR methods shows that additional information is needed for map consideration and more accurate orientation and stride length estimation. The CDF presents the deviation qualitatively and quantitively, for instance, it can be seen that 90% of the calculated points in PDR 2 had an accuracy which was better than 15 m. Electronics 2021, 10, x FOR PEER REVIEW 6 of 17 Finally, PDR is then performed for each detected step. In Equation (1), the coordinates of the current position X  are calculated, depending on the previous position X  , summing with a multiplication of the scalar translation of T (stride length), and heading matrix R. X  =X  + T × R (1 ) The two PDR realizations led to different trajectories (see Figure 3). The trajectories were both subject to significant orientation drift and scale errors caused by incorrect stride length estimation. In order to evaluate and track the approaches quality on-board, a cumulative distribution function (CDF) plot was placed next to each of the estimated trajectories. The CDFs were calculated by fitting a normal distribution to distances between the ground truth step coordinates and the estimated one, i.e., errors. In other words, the error is defined as the nominal value minus the estimated value. If there was a contrast between the pedometer and the actual number of the steps, a simple interpolation reduced or added a few step positions. The evaluation of the PDR methods shows that additional information is needed for map consideration and more accurate orientation and stride length estimation. The CDF presents the deviation qualitatively and quantitively, for instance, it can be seen that 90% of the calculated points in PDR 2 had an accuracy which was better than 15 m. (a) (b) Figure 3. Three different PDR realizations, named PDR 1 (the k-based stride length and Madgwick-based heading), PDR 2 (fixed stride length and leveled gyroscope heading), and PDR mix (the k-based stride length and leveled gyroscope heading): (a) Trajectories shown on the map; and (b) Error vs. CDF probability of the methods. 4.2. 5G-Based Positioning The deployment of cellular systems, from the First Generation (1G) to the current (Fifth Generation, 5G, or New Radio, NR) stage, has a long history. The Third Generation Partnership Project (3GPP) releases have attempted to answer the discovered requirements. The 3GPP unites seven standard telecommunications development organizations. Their members produce reports and specifications that define 3GPP technologies, including radio access, core network, and service capabilities [39]. After a while, the demand on cellular networks came not only from mobile phone providers, but also wider industry. Furthermore, 4G systems have incorporated significant updates addressing markets different than traditional mobile phone businesses [40]; for instance, the device-to-device (D2D) paradigm allows devices to directly communicate with each other using a local wireless channel. D2D technology laid the foundations for the vehicle-to-everything (V2X) technology introduced in 3GPP Release 14 [41]. However, the 5G network is the first cellular network that was not designed with a sole focus on mobile phones. Toward the latter part of the 2010s, the last-generation networks began to reach their limits. The reasons for this were manifold, including a tremendous growth in smartphone penetration and an increase in bandwidth-hungry applications, as well as the large numFigure 3. Three different PDR realizations, named PDR 1 (the k-based stride length and Madgwickbased heading), PDR 2 (fixed stride length and leveled gyroscope heading), and PDR mix (the k-based stride length and leveled gyroscope heading): ( a ) Trajectories shown on the map; and ( b ) Error vs. CDF probability of the methods. 4.2. 5G-Based Positioning The deployment of cellular systems, from the First Generation (1G) to the current (Fifth Generation, 5G, or New Radio, NR) stage, has a long history. The Third Generation Partnership Project (3GPP) releases have attempted to answer the discovered requirements. The 3GPP unites seven standard telecommunications development organizations. Their members produce reports and specifications that define 3GPP technologies, including radio access, core network, and service capabilities [ 39 ]. After a while, the demand on cellular networks came not only from mobile phone providers, but also wider industry. Furthermore, 4G systems have incorporated significant updates addressing markets different than traditional mobile phone businesses [ 40 ]; for instance, the device-to-device (D2D) paradigm allows devices to directly communicate with each other using a local wireless channel. D2D technology laid the foundations for the vehicle-to-everything (V2X) technology introduced in 3GPP Release 14 [ 41 ]. However, the 5G network is the first cellular network that was not designed with a sole focus on mobile phones. Toward the latter part of the 2010s, the last-generation networks began to reach their limits. The reasons for this were manifold, including a tremendous growth in smartphone penetration and an increase in bandwidth-hungry applications, as well as the large number of connected devices. All of these new requirements were gathered and the 3GPP initiated the definition of the 5G technology. The 5G NR requirements have been structured under three main categories [40]: • Enhanced mobile broadband (eMBB), the aim of which is to provide wireless connectivity with very high bandwidth. • Massive machine type communications (mMTC), providing connectivity to a large number of IoT devices, such as smart meters, watches, or wearables. mMTC requires very large cell and network capacities. • Ultra-reliable low latency communications (URLLC), targeted at providing low latency, robust communication links for V2X, remote surgery, and other safety-critical applications. Electronics 2021,10, 397 7 of 17 The 5G network also includes other devices and use-cases, referred to as industry verticals. Another significant difference, with respect to previous cellular networks, is the use of higher frequencies in the millimeter-wave spectrum, starting at 24 GHz. Important features have also come from 5G, such as massive multiple-input and multiple-output (MIMO), beamforming, cloud computing, and network virtualization. All of these features, introduced in Releases 16–17 [ 42 , 43 ], help to increase the scalability and modularity of the network and aid in reaching peak data rates of 20 Gbps with very high user density [40]. 4.2.1. Algorithms and Technologies We have classified the 5G-based positioning approaches based on their geometric bases, analyzing proximity-, distance-, angle-, and time difference-based positioning technologies. In this classification, there is no consideration of the data driven approaches such as the well-known Fingerprinting method. Proximity is the simplest way to determine the location of an object. It is a method which has been long used in cellular networks; for example, in cell ID (CID), Wi-Fi, and Bluetooth positioning systems. Proximity is based on the knowledge that the object to be located is near a reference object whose position is known. Hence, the accuracy of proximity as a positioning method basically depends on the range of the signal used. For instance, dense 5G networks, which are envisioned to range from a few meters up to tens of meters, have shown acceptable accuracy performance when using proximity [ 13 ]. The point representing the position of the mobile terminal can be obtained by geometric calculations, for which the centroid method is the most common [ 44 ]. The weight of each Antenna Node (AN) is proportional to its Received Signal Strength (RSS) value; as a result, the AN with the highest RSS pulls the centroid most strongly to its own position [ 45 ]. The same weighting analysis can be performed using time measurements. However, power measurements are not as accurate as timing measurements in the positioning technologies used in legacy cellular networks (up to LTE). This problem has been solved in 5G NR [ 40 ]. Triangulation is a technique which determines the position of an object by measuring the angles to the object from known points. It relies on Angle of Arrival (AoA) measurements, in which the incoming angle of the received signal is known, such that the receiver can estimate the direction of the transmitter. At least two angle measurements from two different known points are sufficient for 2D localization, by applying trigonometric identification. Although angle-based positioning technologies are not new, due to the beamforming feature, it is with 5G NR that their full potential can be reached. DL-AoD and UL-AoA are two 5G angular positioning technologies that are always linked to a base station beam by 3GPP specification (38.305) [46]. Trilateration is the name given to positioning algorithms referring to a position determined from distance measurements. These are also called range measurement techniques. In Trilateration, the “tri” stands for the (at least) three fixed points that are necessary to determine a 2D position [ 47 ]. In [ 44 ], the use of the least square method to minimize the localization error was discussed, where the absolute distance error between the transmitter and receiver should ideally be zero. Positioning based on trilateration can be performed using several different positioning measurements, such as Time of Arrival (ToA), RSS, and so on. Enhanced Cell ID is a method based on a proximity algorithm for LTE as an enhancement of a timing advance procedure. The aim is to reduce the area of uncertainty of the positioning [ 48 – 50 ]. Alternatively, if the transmitted power is known, the RSS can be used to estimate the distance the signal has travelled, by using particular propagation models that estimate the attenuation of the signal in relation to the travelled distance. This is due to the fact that the power of the transmitted signal decreases with the travelled distance [ 44 ]. In 5G NR, the Round-Trip Time (RTT) to a neighbor cell has been used to define a multi-RTT method. The NR E-CID is also a re-definition of ECID; this time, for 5G networks [40]. Multilateration refers to positioning algorithms based on the difference in distances from the object to two reference points of known location. In particular, they rely on Time Electronics 2021,10, 397 8 of 17 Difference of Arrival (TDoA) measurements. In general, at least two measurements (from three transmitters) are needed to obtain a 2D position, while at least three measurements are required to calculate a 3D position. In 5G NR, DL-TDOA and UL-TDOA are based on the same principle, where the time difference measurement is carried out by the network of base stations [40]. A similar method is the 4G OTDOA. 4.2.2. 5G Simulation Regarding 5G, positioning methods can generate coordinates with a pre-defined resolution, accuracy, and latency. Performance varies in 5G-based technologies and 3GPP releases. For instance, in Release 17, the aim was to generate a horizontal positioning precision of at least 1 meter or higher for industrial use-cases; for example, an accuracy of 0.2 meter is desirable in some indoor use-cases. A positioning latency of less than 100 milliseconds is also desired [ 43 ]. Although a real 5G Campus Network is going to be built in the framework of the Level 5 Indoor Navigation (L5IN) project running at the HafenCity University, an indoor cellular-based positioning simulation was designed to evaluate different antenna placements, network resolutions, qualities, etc. We carried out a simulation based on the points discussed above relating to 5G positioning methods and considering the 5G releases. However, simulation points were derived for simplicity in the following experiments, mainly by using the floor plans and reference points with random errors in a range of 5 m. 5. Autonomous Position Estimation In the following section, the most promising approaches for autonomous pedestrian navigation are presented and discussed. For better understanding of the results, the technical details are also explained. 5.1. PDR Particle Filter The autonomous position estimation was initialized by the PDR 1. PDR-based approaches require correction for pedestrian navigation, in the case of longer-lasting position estimations. Therefore, the available components for navigation were considered; the floor plan and the routing graph. These were intelligently combined with the PDR using PF algorithms, which has been used previously in [38]. In this variant, the PDR represents the Propagation process, and the pedometer controls the computation rate. During the estimation step, the floor plan and the routing graph were used as correction parameters. Figure 4graphically represents the processing of the available data to determine corrections. In Figure 4a, the floor plan is used with a Boolean logic principle: If a particle is behind a wall, change its weight to zero (red). The other particles change their weights, based on their orientation with respect to the wall position (grey to black particles). As Figure 4b shows, the weighting principle was only based on the routing edges. The orthogonal offset has been used as the weight defined on the normal distribution in Equation (4). tan ai=yi−ym xi−xm. (2) wi=exph−0.5·(αi−rwall)·R−1 r·(αi−rwall)i(3) wi=exph−0.5·dorthogonal i·R−1 d·dorthogonal ii (4) Equation (2) shows the calculation of the course between actual particle i and weighted mean value of last estimation step. This angle was used to compare the orientation of walls around this particle. The grey-scaled color of particles in Figure 4a shows the weights based on these comparisons to nearest wall included in a probability density function which represents a normal distribution Equation (3). Equation (4) shows here the probability density function, which represents a normal distribution. A longer distance to routing Electronics 2021,10, 397 9 of 17 edge produced a lower weight for this particle. The weighted mean value out of all particle moved more to routing edge, based on the idea, that this was the typical area of walking. At the end, these corrections generate a trajectory which is positioned more near routing edges, which are typically oriented in the mean axis of corridors. Electronics 2021, 10, x FOR PEER REVIEW 9 of 17 based on the routing edges. The orthogonal offset has been used as the weight defined on the normal distribution in Equation (4). (a) (b) Figure 4. Principles of support: (a) floor plan—red particles are behind wall and are weighted zero; and (b) routing graph—the orthogonal offset of a particle to the nearest routing edge is used to calculate its weight [21]. tana  =y  −y  x  −x  . (2) w  =exp [−0.5∙(α  −r  )∙ R   ∙(α  −r  )] (3) w  =exp [−0.5 ∙ d   ∙ R   ∙d  ] (4) Equation (2) shows the calculation of the course between actual particle i and weighted mean value of last estimation step. This angle was used to compare the orientation of walls around this particle. The grey-scaled color of particles in Figure 4a shows the weights based on these comparisons to nearest wall included in a probability density function which represents a normal distribution Equation (3). Equation (4) shows here the probability density function, which represents a normal distribution. A longer distance to routing edge produced a lower weight for this particle. The weighted mean value out of all particle moved more to routing edge, based on the idea, that this was the typical area of walking. At the end, these corrections generate a trajectory which is positioned more near routing edges, which are typically oriented in the mean axis of corridors. Using the approach of map support on the test data yielded a significantly different trajectory, compared to the PDR trajectory. However, the approach is already clearly less effective. Without further correction using the routing graph, the filter can change the room after a certain time, depending on the noise parameters. 5.2. Topological Approach In the topological approach first presented by [38], the PF is replaced by a state detection approach. A position estimation works only based on the routing graph. The idea here is that the change of priority of orientation estimation makes the PDR more robust for longer durations of navigation. An additional advantage compared to filter approaches is the clear state detection without the influence of randomly distributed parameter repeatability. A further advantage is the requirement of less computing power, compared to a PF. The topological approach works with four state queries. These different states are necessary to handle the different error situations based on PDR influences by step length error and drift error of calculated rotation angle. The basis is a comparison of the currently calculated direction using gyroscope data and the orientation of surrounding routing edges. A selection of the appropriate routing edge is applied. This process is repeated for each step detection. Thus, the PDR runs along the most probable routing edges. One disadvantage is that it cannot assume positions freely in space. Figure 4. Principles of support: ( a ) floor plan—red particles are behind wall and are weighted zero; and ( b ) routing graph—the orthogonal offset of a particle to the nearest routing edge is used to calculate its weight [21]. Using the approach of map support on the test data yielded a significantly different trajectory, compared to the PDR trajectory. However, the approach is already clearly less effective. Without further correction using the routing graph, the filter can change the room after a certain time, depending on the noise parameters. 5.2. Topological Approach In the topological approach first presented by [ 38 ], the PF is replaced by a state detection approach. A position estimation works only based on the routing graph. The idea here is that the change of priority of orientation estimation makes the PDR more robust for longer durations of navigation. An additional advantage compared to filter approaches is the clear state detection without the influence of randomly distributed parameter repeatability. A further advantage is the requirement of less computing power, compared to a PF. The topological approach works with four state queries. These different states are necessary to handle the different error situations based on PDR influences by step length error and drift error of calculated rotation angle. The basis is a comparison of the currently calculated direction using gyroscope data and the orientation of surrounding routing edges. A selection of the appropriate routing edge is applied. This process is repeated for each step detection. Thus, the PDR runs along the most probable routing edges. One disadvantage is that it cannot assume positions freely in space. For the test dataset was used for this approach, see Figure 5. This approach works well if the trajectory follows a clear and simple routing graph. This was not the case here and the algorithm failed after about 60% of the run (Figure 5). The trajectory could not match the routing graph and the calculate its path. Uncertainties in positioning, therefore, increase strongly if the traversed path is not well-represented by the routing edges, or if it deviates strongly. Electronics 2021,10, 397 16 of 17 8. Yang, J.; Cao, Z.; Zhang, Q. A fast and robust local descriptor for 3D point cloud registration. Inf. Sci. 2016 ,346–347, 163–179. [CrossRef] 9. Dube, R.; Dugas, D.; Stumm, E.; Nieto, J.; Siegwart, R.; Cadena, C. SegMatch: Segment based place recognition in 3D point clouds. In Proceedings of the 2017 IEEE International Conference on Robotics and Automation (ICRA), Singapore, 29 May–3 June 2017; pp. 5266–5272. 10. Hess, W.; Kohler, D.; Rapp, H.; Andor, D. Real-time loop closure in 2D LIDAR SLAM. In Proceedings of the 2016 IEEE International Conference on Robotics and Automation (ICRA), Stockholm, Sweden, 16–20 May 2016; Volume 2016, pp. 1271–1278. 11. Grisetti, G.; Tipaldi, G.D.; Stachniss, C.; Burgard, W.; Nardi, D. Fast and accurate SLAM with Rao–Blackwellized particle filters. Robot. Auton. Syst. 2007,55, 30–38. [CrossRef] 12. Montemerlo, M.; Thrun, S.; Koller, D.; Wegbreit, B. FastSLAM: A factored solution to the simultaneous localization and mapping problem. In Proceedings of the AAAI National Conference on Artificial Intelligence/IAAI, Edmonton, AB, Canada, 28 July–2 August 2002; p. 593598. 13. Koivisto, M.; Hakkarainen, A.; Costa, M.; Kela, P.; Leppänen, K.; Valkama, M. High-Efficiency Device Positioning and Loca-tionAware Communications in Dense 5G Networks. IEEE Commun. Mag. 2017,55, 188–195. [CrossRef] 14. Xiao, Z.; Wen, H.; Markham, A.; Trigoni, N. Robust Indoor Positioning with Lifelong Learning. IEEE J. Sel. Areas Commun. 2015 , 33, 2287–2301. [CrossRef] 15. Mizell, D.W. Using gravity to estimate accelerometer orientation. In Proceedings of the Seventh IEEE International Symposium on Wearable Computers, White Plains, NY, USA, 21–23 October 2003; Institute of Electrical and Electronics Engineers (IEEE): Los Alamitos, CA, USA, 2004; pp. 252–253. 16. Kunze, K.; Lukowicz, P.; Partridge, K.; Begole, B. Which way am I facing: Inferring horizontal device orientation from an ac-celerometer signal. In Proceedings of the 2009 International Symposium on Wearable Computers, Linz, Austria, 4–7 September 2009; pp. 149–150. 17. All About the Sensors that Make the iPhone so Cool. Available online: https://www.lifewire.com/sensors-that-make-iphone-socool-2000370 (accessed on 5 November 2020). 18. Woodman, O.; Harle, R. Pedestrian Localisation for Indoor Environments. Ph.D. Thesis, University of Cambridge, Cambridge, MA, USA, 2010. 19. Constandache, I.; Choudhury, R.R.; Rhee, I. Towards Mobile Phone Localization without War-Driving. In Proceedings of the 2010 IEEE INFOCOM, San Diego, CA, USA, 15–19 December 2010; pp. 1–9. 20. Google Maps’ Biggest Moments over the Past 15 Years. Available online: https://blog.google/products/maps/look-back-15 -years-mapping-world/ (accessed on 5 November 2020). 21. Willemsen, T. Fusionsalgorithmus zur Autonomen Positionsschätzung im Gebäude, Basierend auf MEMS-Inertialsensoren im Smartphone. Ph.D. Thesis, HafenCity University, Hamburg, Germany, 2016. 22. Frassl, M. Harnessing Geomagnetic Field Disturbances for Ubiquitous Navigation. Ph.D. Thesis, Ulm University, Ulm, Germany, 2018. 23. Lukianto, C.; Sternberg, H. STEPPING—Smartphone-Based Portable Pedestrian Indoor Navigation. Arch. Fotogram. Kartogr. Teledetekcji 2011,22, 311–323. 24. Brajdic, A.; Harle, R.K. Walk detection and step counting on unconstrained smartphones. In Proceedings of the 2013 ACM International Joint Conference on Pervasive and Ubiquitous Computing, Zurich, Switzerland, 8–12 September 2013; Association for Computing Machinery (ACM): New York, NY, USA, 2013; pp. 225–234. 25. Li, F.; Zhao, C.; Ding, G.; Gong, J.; Liu, C.; Zhao, F. A Reliable and Accurate Indoor Localization Method Using Phone Inertial Sensors. In Proceedings of the 2012 ACM Conference on Ubiquitous Computing, Pittsburgh, PA, USA, 5–8 September 2012; pp. 421–430. 26. Chen, C.; Lu, X.; Markham, A.; Trigoni, N. IoNet: Learning to cure the curse of drift in inertial odometry. In Proceedings of the AAAI Conference on Artificial Intelligence, New Orleans, LA, USA, 2–7 February 2018; Volume 32, pp. 6468–6476. 27. Yan, H.; Shan, Q.; Furukawa, Y. RIDI: Robust IMU Double Integration. In Proceedings of the European Conference on Computer Vision (ECCV), Munich, Germany, 8–14 September 2018; pp. 621–636. 28. Chen, C.; Zhao, P.; Lu, C.X.; Wang, W.; Markham, A.; Trigoni, N. Deep-Learning-Based Pedestrian Inertial Navigation: Methods, Data Set, and On-Device Inference. IEEE Internet Things J. 2020,7, 4431–4441. [CrossRef] 29. Cortés, S.; Solin, A.; Rahtu, E.; Kannala, J. ADVIO: An Authentic Dataset for Visual-Inertial Odometry. Min. Data Financ. Appl. 2018,11214 LNCS, 425–440. [CrossRef] 30. Xiao, Z. Robust Indoor Positioning with Lifelong Learning. Ph.D. Thesis, University of Oxford, Oxford, UK, 2014. 31. Kuang, J.; Niu, X.; Chen, X. Robust Pedestrian Dead Reckoning Based on MEMS-IMU for Smartphones. Sensors 2018 ,18, 1391. [CrossRef] [PubMed] 32. Hoseinitabatabaei, S.A.; Gluhak, A.; Tafazolli, R.; Headley, W. Design, realization, and evaluation of uDirect—An approach for pervasive observation of user facing direction on mobile phones. IEEE Trans. Mob. Comput. 2014,13, 1981–1994. [CrossRef] 33. Roy, N.; Wang, H.; Choudhury, R.R. I am a Smartphone and I can Tell my User’ s Walking Direction. In Proceedings of the 12th Annual International Conference on Mobile Systems, Applications, and Services, Bretton Woods, NH, USA, 16–19 June 2014; pp. 329–342. Electronics 2021,10, 397 17 of 17 34. Samsung Galaxy S10 5G. Available online: https://www.samsung.com/de/smartphones/galaxy-s10/galaxy-s10-5g/ (accessed on 14 December 2020). 35. Sensor Log-Apps on Google Play. Available online: https://play.google.com/store/apps/details?id=com.hfalan.activitylog&hl= en (accessed on 14 December 2020). 36. Invensense MPU-9250 | TDK. Available online: https://invensense.tdk.com/products/motion-tracking/9-axis/mpu-9250/ (accessed on 11 December 2020). 37. Madgwick, S.O.H. An Efficient Orientation Filter for Inertial and Inertial/Magnetic Sensor Arrays. 2010. Available online: https://www.x-io.co.uk/res/doc/madgwick_internal_report.pdf (accessed on 3 April 2010). 38. Willemsen, T.; Keller, F.; Sternberg, H. A topological approach with MEMS in smartphones based on routing-graph. In Proceedings of the 2015 International Conference on Indoor Positioning and Indoor Navigation (IPIN), Banff, AB, Canada, 13–16 October 2015; pp. 1–6. [CrossRef] 39. About 3GPP. Available online: https://www.3gpp.org/about-3gpp/about-3gpp (accessed on 5 November 2020). 40. García, A.; Maier, S.; Philips, A. Location-Based Services in Cellular Networks: From GSM to 5G NR; Artech House: London, UK, 2020. 41. Release 14. Available online: https://www.3gpp.org/release-14 (accessed on 5 November 2020). 42. Release 16. Available online: https://www.3gpp.org/release-16 (accessed on 5 November 2020). 43. Release 17. Available online: https://www.3gpp.org/release-17 (accessed on 9 December 2020). 44. Alyafawi, I. Real-Time Localization using Software Defined Radio. Ph.D. Thesis, Universität Bern, Bern, Switzerland, 2015. 45. Celik, G.; Celebi, H.; Tuna, G. A novel RSRP-based E-CID positioning for LTE networks. In Proceedings of the 13th International Wireless Communications and Mobile Computing Conference (IWCMC), Valencia, Spain, 26–30 June 2017; pp. 1689–1692. [CrossRef] 46. 3GPP Specification Series: 38series. Available online: https://www.3gpp.org/DynaReport/38-series.htm (accessed on 10 December 2020). 47. Correa, A.; Barcelo, M.; Morell, A.; Vicario, J.L. A Review of Pedestrian Indoor Positioning Systems for Mass Market Applications. Sensors 2017,17, 1927. [CrossRef] [PubMed] 48. Mike, T.; Ewald, Z. LTE Location Based Technology Introduction. Available online: https://www.rohde-schwarz.com/kr/ applications/white-paper_230854-122561.html (accessed on 29 May 2015). 49. Campos, R.S. Evolution of Positioning Techniques in Cellular Networks, from 2G to 4G. Wirel. Commun. Mob. Comput. 2017 ,2017, 1–17. [CrossRef] 50. Borkowski, J.; Niemelä, J.; Lempiäinen, J. Enhanced performance of Cell ID+RTT by implementing forced soft handover algo-rithm. In Proceedings of the IEEE 60th Vehicular Technology Conference, Los Angeles, CA, USA, 26–29 September 2004; Volume 60, pp. 3545–3549.