Unmanned Aerial Vehicle Navigation with Event Cameras

Unmanned Aerial Vehicle navigation with event cameras Abstract For proper navigation of Unmanned Aerial Vehicles (UAVs), it is necessary to know their position in real-time to ensure safe navigation. Determining position in outdoor spaces is quite well solved. On the other hand, in indoor spaces, existing solutions are either imprecise or excessively costly. In this paper, the 3D localization problem is addressed in the context of UAV navigation. The main purpose of this work is to develop and evaluate a robust real-time localization scheme using exclusively the information from an embedded Event Camera and an IMU (Inertial Measurement Unit). Deep learning techniques and robust computer vision algorithms are implemented together to accurately compute the UAV pose, leveraging the strengths of well-established visual-inertial odometry algorithms and the intrinsic advantages of Event Cameras, such as high dynamic range and absence of motion blur. Throughout this study, state-of-the-art techniques are selected, refined, implemented, and evaluated. The proposed system demonstrated good performance and acceptable precision specially in situation with abrupt lighting changes. Keywords: Event camera, Unmanned Aerial Vehicle, Deep Learning, Visual-Inertial Odometry, Sensor Fusion, Pose Estimation. 1. Introduction Within robotics, the quality of navigation depends on both the embedded sensors and the type of environment in which the robot operates. Embedded sensors enable us to perceive the surrounding world, providing crucial data for accurately estimating the robot position and orientation. The environment type in which navigation occurs significantly influences the amount of useful information inferred from sensor measurements. Thus, appropriately contextualizing existing navigation solutions is essential to introduce the motivation for this project. Preprint submitted to To decide This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4861406 Preprint not peer reviewed (a) GNSS navigation scheme. (b) LiDAR localization and mapping. Figure 1: Examples of existing solutions for UAV navigation. The navigation of Unmanned Aerial Vehicles (UAVs) in outdoor spaces is typically precise enough for safe operation, thanks to robust satellite coverage and the capability to identify magnetic north. The key components of this navigation system include the Global Navigation Satellite System (GNSS), Inertial Measurement Units (IMUs), and magnetometers. These elements collectively contribute to the accurate estimation of the drone pose, ensuring safe navigation. Research efforts, such as the one detailed in [1], leverage these technologies for applications such as package delivery with drones. An illustrative overview of its functioning is provided in 1a. However, the situation changes radically when navigation is carried out in an indoor environment where GNSS coverage is either unavailable or of low quality. These scenarios are commonly referred to as ’GNSS-denied environments’. In these cases, navigation is typically addressed using a diverse array of techniques and technologies. Other works have studied localization and mapping techniques within these spaces using mainly cameras and inertial sensors. These works provided valuable generic algorithms [2? ? ? ] and application-specific research [3?]. These systems provide a cost-effective solution for the real-time localization problem with enough precision for many applications [4]. However, there are many challenging scenarios in which the performance of these algorithms drops significantly, especially in complex and dynamic environments [5]. One of the widely used sensors is based on Light Detection and Ranging (LiDAR) technology. These sensors are capable of reconstructing the environment by detecting a cloud of points. In [6], 2D LiDAR sensors were employed to support terrain robot navigation. In [7], 3D LiDAR sensors were 2 This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4861406 Preprint not peer reviewed utilized to map and inspect infrastructures with UAVs. Figure 1b shows an example of environmental reconstruction and localization using a LiDAR sensor. However, the technologies mentioned above present some issues. LiDAR sensors are heavy and expensive, leading to a reduction in UAV flight time and an increase in the overall solution cost. On the other hand, conventional cameras encounter challenges with lightness and fast movements, resulting in the well-known problem of motion blur. Event Cameras (EC) are novel bio-inspired sensors invented to address issues encountered with conventional cameras. ECs transmit independent pixel information as soon as their intensity change exceeds an upper or lower threshold. Events, or pixel change notifications, occur asynchronously per pixel, with a microsecond resolution, resulting in minimal latency. Unmanned Aerial Vehicles (UAVs), particularly multi-rotors, are increasingly popular in many tasks nowadays, as they offer the potential to mitigate human risks and effectively address repetitive tasks. These tasks span from surveillance or warehouse inventory to operations in hazardous environments, such as infrastructure inspection or search and rescue applications [8]. Throughout this paper, an Event Camera (EC)-based localization method is proposed to address autonomous drone navigation in indoor environments. In this context, the use of ECs appears to be particularly useful, thanks to their inherent and distinctive features that can be effectively leveraged in visually rich surroundings. The aim is to introduce and evaluate a system that assesses the feasibility of implementing this technology. 2. Previous work Due to the nascent stage of ECs, they have only recently garnered attention in research circles. In the literature, some studies propose algorithms exclusively utilizing events, while others integrate events as components of a broader system. A brief analysis of the main research trends with event cameras related to the target localization system will be provided next. 2.1. Visual feature tracking This refers to a fundamental task commonly used in the development of more complex vision algorithms, such as object tracking or visual odometry. Works such as [9] or [10] adapt classic algorithms to identify simple patterns in the event paradigm. The low latency of event cameras and their ability 3 This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4861406 Preprint not peer reviewed to react to motion make them ideal sensors for obtaining information in dynamic situations. Unlike standard cameras, ECs enable smooth object tracking without being affected by the blind time between frames typical of standard cameras. 2.2. Video reconstruction Driven by the capability of events to capture all variations of light intensity in a scene, it becomes possible to reconstruct frames from a temporal window of events. This task enables the connection of this novel technology with decades of research on algorithms for traditional cameras. Early attempts to solve this task using classical approaches, such as [11], often encountered limitations. It was not until the introduction of deep learning techniques that major improvements were made in the field. A groundbreaking example is the E2VID architecture [12], which achieved high-quality image reconstruction across various scenarios using recurrent convolutional neural networks. Subsequently, the widespread adoption of deep learning techniques led to numerous works aiming to improve either inference time or image reconstruction quality. Notable examples include the FireNet architecture [13], which maintains state-of-the-art reconstruction quality with substantially reduced inference time, and the introduction of ET-Net [14] leveraging transformers for enhancing the reconstruction quality. Additionally, [15] developed HyperE2VID, which incorporates hypernetworks to dynamically modify the net parameters for increased adaptability. 2.3. Simultaneous localization and mapping This refers to the computational challenge of concurrently constructing a map of an unknown environment while accurately estimating an agent pose within it. Thus, these algorithms are a solid option to solve the problem of UAV localization addressed in this work. Early efforts, exemplified by [16], achieved continuous camera pose tracking using pre-generated depth maps. Subsequently, more complex sensor fusion approaches emerged, integrating ECs with Inertial Measurement Units (IMUs) and conventional cameras to estimate the target data through advanced techniques, as in the work [17]. Recent works such as [18] have probabilistically integrated frames and events to achieve this objective. Moreover, the exploration of stereo pairs of event cameras, as investigated in [19], has shown promising results in depth estimation and more precise localization compared to monocular systems. 4 This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4861406 Preprint not peer reviewed In conclusion, while significant advances have been made, these works show that there are still opportunities to enhance the robustness and efficiency of these systems. 2.4. Embedded event-based vision for UAV navigation Several works have embedded an event camera in a UAV to address the problem of obstacle detection and avoidance, leveraging its low latency to successfully avoid dynamic objects. An example of this line of investigation is presented in [20], where a UAV is able to detect and avoid fast-moving obstacles using only event-driven sensing and onboard computation. In this study, temporal windows of events are ego-motion compensated to directly detect obstacles. Another example is [21], where the same goal is achieved using a Spiking Neural Network (SNN) to identify dynamic obstacles in the event stream. In this case, a depth camera is also used to make the proper decision to maneuver and avoid the obstacle. Moreover, [22] leverages event-driven data and SNNs to develop a UAV vision-based position control system implemented on a neuromorphic chip. In this context, the system achieves high-speed UAV control while maintaining low latency and low power consumption. Another interesting project is [23], where a UAV can be teleoperated with minimal delay through visual gestures using both a conventional camera and an event camera. Some of these works outperform equivalent standard camera algorithms, but they also encounter limitations regarding their application in various real-world scenarios [24]. This work, emphasis is placed on addressing the localization challenge within the domain of UAV navigation. By employing an intermediate representation, our aim is to address this problem in a more robust and versatile manner. 3. Material and Methods In order to understand the essence of this work, it is essential to initially introduce the physical systems which will be used to implement and validate the proposal of localization algorithm. 3.1. Event camera This work has used the IniVation DVXplorer camera, an event camera with a resolution of 640x480 pixels which does not generate conventional frames. Events are produced with a time resolution of 200 microseconds, 5 This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4861406 Preprint not peer reviewed Figure 2: DVXplorer event camera. Figure 3: UAV platform. and it has a latency of less than 1 ms, as stated in its datasheet 1. The camera is also equipped with a lens, enabling manual adjustments for focus, optical zoom, and aperture. Refer to Figure 2 to observe the DVXplorer camera. 3.2. Testing environment To evaluate the localization system, it is necessary to reliably know the UAV pose during testing. Therefore, the testing facility used is the multi- UAS laboratory of **anonymized**, a 15 x 15 x 5 m room equipped with 20 VICON cameras. These cameras emit infrared light at 500 Hz, which reflects off small spheres placed on the multi-rotor body. These cameras capture the reflected infrared light from different angles, enabling them to precisely triangulate the position of each sphere. This allows the VICON system to recognize the UAV and accurately compute its pose, which will serve as ground-truth. A VICON camera can be seen in Figure 5. Since a camera-dependent odometry algorithm is evaluated in this work, it is essential to operate in a rich visual environment. This is the reason why various shelves and objects have been strategically placed in the room, as depicted in Figure 4. 3.3. UAV platform The UAV platform is built around the DJI F550 hexarotor frame, as depicted in Figure 3. The autopilot, a PixHawk Orange Cube running Ardupi- 1https://inivation.com/wp-content/uploads/2021/08/2021-08-iniVation-devices- Specifications.pdf 6 This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4861406 Preprint not peer reviewed Figure 4: **anonymized** multi-UAS laboratory. Figure 5: VICON camera. lot, is centrally mounted on the drone center of mass. Alongside the autopilot is an onboard computer (Nvidia Jetson Xavier NX) which is responsible for executing the algorithms required for the operations. A wireless communication system between the onboard PC and the development PC is also integrated, allowing convenient data capture and system deployment. Notably, the event camera is securely affixed to the drone thanks to a slightly tilted mount, which is also equipped with rubber dampers to mitigate the highfrequency vibrations during the flight. Furthermore, due to the sensitivity of the selected event camera to high-frequency infrared light emitted by the VICON cameras, it is necessary to affix an infrared filter lens to the camera objective. This prevents the introduction of significant noisy interference. 3.4. Proposed localization system Based on the review of the existing state-of-the-art literature, the integration of conventional computer vision techniques with event-based processing using an intermediate deep learning architecture is proposed. The intention is to generate synthetic frames that serve as a bridge, connecting the advantages of event cameras with the robust and mature algorithms of traditional computer vision. The primary advantage of these reconstructed frames lies in their potential to achieve a significantly high dynamic range and remain free of motion blur, in contrast to conventional frames. These attributes are enabled by the stream of events, which encapsulates all visual variations with high temporal resolution. Therefore, this work aims to address the localization problem by employing an artificial neural network to generate reconstructed frames, followed by a robust visual-inertial odometry algorithm to obtain the UAV pose by 7 This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4861406 Preprint not peer reviewed Figure 6: Proposed localization system with intermediate representations. leveraging these synthetic frames and the available inertial measurements. The employed pipeline is represented in Figure 6. 3.5. Image reconstruction As discussed in the state-of-the-art review, image reconstruction from events remained a challenge until the emergence of deep learning techniques. The primary obstacles to successful reconstruction included the cumulative impact of subtle structural disparities among photoreceptors, variations dependent on ambient conditions, and the inherent uncertainty of initial pixel values. The selected architecture for this module is FireNet [13]. As highlighted by the authors, FireNet demonstrates the ability to achieve reconstruction quality approaching the state-of-the-art while employing a simplified neural architecture. The aim is to deploy this architecture in real-time scenarios, underscoring the importance of faster inference to attain high-resolution and low-latency odometry in subsequent modules. The original network weights provided by the authors were obtained through a supervised learning scheme. In this work is employed this pretrained model as the initial framework for fine-tuning the network with application-specific generated datasets. The objective of the fine-tuning is to adapt the default weights to the specific camera and environment of this work, aiming to optimize the reconstruction quality of the network. To accomplish this, the ESIM simulator [25] is utilized to generate the desired pairs of event packages along with their associated ground-truth frames. Employing its planar rendering mode, a custom event camera is simulated 8 This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4861406 Preprint not peer reviewed Figure 7: generating data with ESIM for fine-tuning process: environment photo used as mural (left) and generated pair of event package and frame (right). Figure 8: sequence of synchronized event packages and frames. using the intrinsic parameters acquired from a prior calibration process. The simulated event camera moves three-dimensionally around an artificial mural depicting the environment, thereby generating the requisite data. In Figure 7, the environment photo used and an example of the generated data can be observed. It is noteworthy that the same resolution, distortion model, focal length, and field of view as that of our DVXplorer event camera are simulated. Random camera movements are simulated over a period of 70 seconds, resulting in a lengthy sequence of events and frames. To create a structured dataset, all events occurring between two consecutive frames are packaged, associating each package with its corresponding frame as illustrated in Figure 8. To generate the final dataset for fine-tuning purposes, 20 sequences are randomly sampled, each comprising 10 of these pairs. Furthermore, Gaussian noise is added to each event package to enhance the network capability for reconstruction in the presence of noise. Hence, these sequences are employed to conduct several training epochs 9 This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4861406 Preprint not peer reviewed and odometry. In a regular camera system, abrupt changes in lighting could lead to a loss of tracked feature points due to their sudden alteration in appearance. max mean std APE (m) 1.7077 0.4147 0.1707 RPE (m) 1.4283 0.0120 0.0298 Table 4: Experiment 3. Error measurements. In the same way, the graphs can be seen in Figure 15c, and their values gathered in Table 4. A short peak in the plots is observed, representing a temporary loss of UAV visibility in the VICON system. The average values of RPE and APE remain similar to previous cases. Notably, in Figure 14, one of the instants when the light is turned off is observed. The sensor captures numerous negative events (depicted in blue) due to the sudden decrease in intensity. However, the image reconstruction remains unaffected, and the characteristic points remain stable. 5. Discussion and future work In this article, a localization scheme based on visual-inertial odometry over the reconstructed image from an event camera was presented. The image reconstruction served as a crucial intermediate step, forming the core of the proposed system. The main limitations encountered arise from being constrained to estimations made with a single camera at a maximum frequency determined by the inference capabilities of the equipment. However, despite these challenges, the system demonstrated good performance and acceptable precision, particularly notable in the third experiment, where abrupt lighting changes had minimal impact. Future work could involve integrating additional sensor information to guide the initialization process and improve performance. For instance, setting up an event camera stereo pair and evaluating performance in this configuration could provide valuable insights. At a lower level, advanced fine-tuning processes with more varied data generation could further increase the quality of reconstructions. Additionally, as computational technology advances, improvements in hardware capacity would enable higher temporal resolution and real-time deployment of more demanding reconstruction architectures, potentially leading to obtaining higher-quality results. 16 This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4861406 Preprint not peer reviewed (a) Experiment 1. (b) Experiment 2. (c) Experiment 3. Figure 15: Estimated trajectory (left), APE (middle), RPE (right). 17 This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4861406 Preprint not peer reviewed Lastly, it is worth mentioning that developing neural network architectures to directly compute events-inertial odometry is another promising avenue to explore. The objective would be to achieve robust and efficient results without the need for the intermediate reconstruction process. This approach could potentially streamline the localization pipeline and improve overall system performance. References [1] A. Patrik, G. Utama, A. A. S. Gunawan, A. Chowanda, J. S. Suroso, R. Shofiyanti, W. Budiharto, Gnss-based navigation systems of autonomous drone for delivering items, Journal of Big Data 6 (2019) 1–14. [2] T. Qin, P. Li, S. Shen, Vins-mono: A robust and versatile monocular visual-inertial state estimator, IEEE Transactions on Robotics 34 (4) (2018) 1004–1020. [3] G. Balamurugan, J. Valarmathi, V. P. S. Naidu, Survey on uav navigation in gps denied environments, in: 2016 International Conference on Signal Processing, Communication, Power and Embedded System (SCOPES), 2016, pp. 198–204. doi:10.1109/SCOPES.2016.7955787. [4] A. Macario Barros, M. Michel, Y. Moline, G. Corre, F. Carrel, A comprehensive survey of visual slam algorithms, Robotics 11 (1) (2022) 24. [5] K. Minoda, F. Schilling, V. W¨uest, D. Floreano, T. Yairi, Viode: A simulated dataset to address the challenges of visual-inertial odometry in dynamic environments, IEEE Robotics and Automation Letters 6 (2) (2021) 1343–1350. [6] J. Sun, J. Zhao, X. Hu, H. Gao, J. Yu, Autonomous navigation system of indoor mobile robots using 2d lidar, Mathematics 11 (6) (2023). doi: 10.3390/math11061455. [7] R. Caballero, J. Parra, M. ´ A. Trujillo, F. J. P´erez-Grau, A. Viguria, A. Ollero, Aerial robotic solution for detailed inspection of viaducts, Applied Sciences 11 (18) (2021) 8404. [8] H. Shakhatreh, A. H. Sawalmeh, A. Al-Fuqaha, Z. Dou, E. Almaita, I. Khalil, N. S. Othman, A. Khreishah, M. Guizani, Unmanned aerial 18 This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4861406 Preprint not peer reviewed vehicles (uavs): A survey on civil applications and key research challenges, Ieee Access 7 (2019) 48572–48634. [9] T. Delbruck, M. Lang, Robotic goalie with 3ms reaction time at 4using event-based dynamic vision sensor, Frontiers in neuroscience 7 (2013) 223. doi:10.3389/fnins.2013.00223. [10] A. Glover, C. Bartolozzi, Robust visual tracking with a freely-moving event camera, in: 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2017, pp. 3769–3776. doi: 10.1109/IROS.2017.8206226. [11] M. Cook, L. Gugelmann, F. Jug, C. Krautz, A. Steger, Interacting maps for fast visual interpretation, in: The 2011 International Joint Conference on Neural Networks, IEEE, 2011, pp. 770–776. [12] H. Rebecq, R. Ranftl, V. Koltun, D. Scaramuzza, High speed and high dynamic range video with an event camera, IEEE transactions on pattern analysis and machine intelligence 43 (6) (2019) 1964–1980. [13] C. Scheerlinck, H. Rebecq, D. Gehrig, N. Barnes, R. Mahony, D. Scaramuzza, Fast image reconstruction with an event camera, in: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2020, pp. 156–163. [14] W. Weng, Y. Zhang, Z. Xiong, Event-based video reconstruction using transformer, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 2563–2572. [15] B. Ercan, O. Eker, C. Saglam, A. Erdem, E. Erdem, Hypere2vid: Improving event-based video reconstruction via hypernetworks, arXiv preprint arXiv:2305.06382 (2023). [16] G. Gallego, J. E. Lund, E. Mueggler, H. Rebecq, T. Delbruck, D. Scaramuzza, Event-based, 6-dof camera tracking from photometric depth maps, IEEE Transactions on Pattern Analysis and Machine Intelligence 40 (10) (2018) 2402–2412. doi:10.1109/TPAMI.2017.2769655. [17] A. R. Vidal, H. Rebecq, T. Horstschaefer, D. Scaramuzza, Ultimate slam? combining events, images, and imu for robust visual slam in hdr 19 This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4861406 Preprint not peer reviewed and high-speed scenarios, IEEE Robotics and Automation Letters 3 (2) (2018) 994–1001. doi:10.1109/LRA.2018.2793357. [18] J. Hidalgo-Carri´o, G. Gallego, D. Scaramuzza, Event-aided direct sparse odometry, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 5781–5790. [19] Y. Zhou, G. Gallego, S. Shen, Event-based stereo visual odometry, IEEE Transactions on Robotics 37 (5) (2021) 1433–1450. [20] D. Falanga, K. Kleber, D. Scaramuzza, Dynamic obstacle avoidance for quadrotors with event cameras, Science Robotics 5 (40) (2020) eaaz9712. [21] X. Zhang, J. Tie, J. Li, Y. Hu, S. Liu, X. Li, Z. Li, X. Yu, J. Zhao, Z. Wan, et al., Dynamic obstacle avoidance for unmanned aerial vehicle using dynamic vision sensor, in: International Conference on Artificial Neural Networks, Springer, 2023, pp. 161–173. [22] A. Vitale, A. Renner, C. Nauer, D. Scaramuzza, Y. Sandamirskaya, Event-driven vision and control for uavs on a neuromorphic chip, in: 2021 IEEE International Conference on Robotics and Automation (ICRA), IEEE, 2021, pp. 103–109. [23] J. P. Rodr´ıguez-G´omez, R. Tapia, A. G. Egu´ıluz, J. R. Mart´ınez-de Dios, A. Ollero, Uav human teleoperation using event-based and frame-based cameras, in: 2021 Aerial Robotic Systems Physically Interacting with the Environment (AIRPHARO), IEEE, 2021, pp. 1–5. [24] M. Ben Miled, Q. Zeng, Y. Liu, Discussion on event-based cameras for dynamic obstacles recognition and detection for uavs in outdoor environments, in: UKRAS22 Conference” Robotics for Unconstrained Environments” Proceedings, EPSRC UK-RAS Network, 2022. [25] H. Rebecq, D. Gehrig, D. Scaramuzza, ESIM: an open event camera simulator, Conf. on Robotics Learning (CoRL) (Oct. 2018). [26] A. I. Mourikis, S. I. Roumeliotis, A multi-state constraint kalman filter for vision-aided inertial navigation, in: Proceedings 2007 IEEE International Conference on Robotics and Automation, 2007, pp. 3565–3572. doi:10.1109/ROBOT.2007.364024. 20 This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4861406 Preprint not peer reviewed [27] P. Geneva, K. Eckenhoff, W. Lee, Y. Yang, G. Huang, Openvins: A research platform for visual-inertial estimation, in: 2020 IEEE International Conference on Robotics and Automation (ICRA), 2020, pp. 4666–4672. doi:10.1109/ICRA40945.2020.9196524. 21 This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4861406 Preprint not peer reviewed Unmanned Aerial Vehicle navigation with event cameras David Tejero-Ruiz1,2, David Sol´ıs-Mart´ın1*, Francisco J. P´erez-Grau2, Juan Gal´an-P´aez1 1*Department of Computer Science and Artificial Intelligence Sevilla University, Seville, Spain, Seville University, Seville, Spain. 2Advanced Center for Aerospace Techonologies (CATEC), Seville, Spain. *Corresponding author(s). E-mail(s): [email protected]; Contributing authors: [email protected];fjp[email protected]; [email protected]; ORCID •David Sol´ıs-Mart´ın: 0000-0002-4625-4554 •Juan Gal´an-P´aez: 0000-0003-0528-9459 Acknowledgments This work has been supported by Proyecto PID2019-109152GB-I00 funded by MCIN/ AEI /10.13039/501100011033 (Agencia Estatal de Investigaci´on), Spain, and SIMAR project, funded by the European Union’s Horizon Europe research and innovation programme under grant agreement No 101070604. 1 This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4861406 Preprint not peer reviewed