Acta Polytechnica CTU Proceedings https://doi.org/10.14311/APP.2024.51.0053 Acta Polytechnica CTU Proceedings 51:53–68, 2024 © 2024 The Author(s). Licensed under a CC-BY 4.0 licence Published by the Czech Technical University in Prague DIGITAL RECONSTRUCTION OF REAL (NEAR) ACCIDENTS OF VULNERABLE ROAD USERS (VRU) IN THE CZECH REPUBLIC – MODERN TECHNOLOGY BRINGS NEW PERSPECTIVES Ondřej Piksaa,∗, Jan Ritterb, Pavel Drbohlavc, Tomáš Petránc a Czech Technical University in Prague, Faculty of Transportation Sciences, Institute of Transport Vehicles, Horská 3, 128 03 Prague 2, Czech Republic b Aurel CZ s.r.o., Road Safety Department, Štěpánská 539/2, 120 00 Prague 2, Czech Republic c Aurel CZ s.r.o., 3D Measurement Department, Štěpánská 539/2, 120 00 Prague 2, Czech Republic ∗ corresponding author: piksa@fd.cvut.cz Abstract. Protecting vulnerable road users (VRUs) remains an issue despite significant advances in the development of integrated safety, autonomous driving and cooperative environments. As part of accident/collision processes, the urban area of municipalities continues to constitute a risk environment with a relatively high frequency of serious injuries (VRU) even at relatively low speeds. In order to understand and increase road safety, it is necessary to analyse the causes of traffic accidents in more detail using video recordings, virtual reality (VR) and automated computer vision methods. The state of the art in the Czech Republic is limited due to the absence of high-quality and extensive data sets. This article constitutes a proof-of-concept solution for the digital reconstruction of selected accident/collision events in 4 cities of the Republic, specifically in Prague, Liberec, Česká Lípa and Jaroměř. It makes it easier to identify, parameterize and virtualize the results of a traffic conflict study and can provide input data for the testing and validation of sensors, interaction models of road users’ behaviour, and preparation of scenarios for a simulated environment; ultimately it reinforces prevention efforts. Keywords: Vulnerable road user (VRU), accident/collision events, road safety, virtual reality (VR), photogrammetry, integrated rescue system (IRS), Czech Republic. 1. Introduction Based on long-term cooperation in the development of cars at Škoda Auto, involvement in preventive IRS programmes and educational activities for primary and secondary schools, we have decided to better un- derstand the occurrence of traffic accidents involving vulnerable road users (VRU) to map the state of the art and subsequently to create a proof-of-concept solu- tion for its detailed analysis. Using the tools of modern technology, such as virtual reality, video analysis, pho- togrammetry and laser scanning, we obtain a unique insight that helps us understand and evaluate better the key parameters of these adverse events in the con- text of space and time. Their prevention/elimination remains the key goal in the near future, especially with the contribution of better proposals for sensory functions, both of vehicles and cooperative transport environments, as well as test scenarios as such and more intensive/better preventive and educational ac- tivities. Data collection has been handled in the project using public camera infrastructure, on-board vehicle cameras, 3D laser scanners, cameras and UAVs. The next step is the digital reconstruction itself, with the final export to virtual reality and the subsequent use of HMD technology to simulate and experience a real accident in the digital world. In addition to the methodological procedure of the solution, we have created our own software application serving primarily the IRS units. We have validated the procedures suggested by use and their results by comparing them with the commercial Virtual CRASH solution. The Police of the Czech Republic has successfully adopted our insights into its practice and nationwide trainings of its members have been held to use the created tools. In cooperation with the Czech Police and the Vision 0 platform, preventive actions using VR simulations based on real traffic accidents are still ongoing. In the future, we would like to upgrade our approach from the proof of concept stage to a more robust application in practice. In the context of the entire work, it is also important to mention the growing need to collect data from accidents that not only have health consequences, but also cause material damage [1]. Both WHO and EU statistics [2, 3] show that vulnerable road users (VRUs) account for more than half of all road accident deaths. In the EU, VRU collisions with a vehicle constitute almost 70% of the total number of fatal accidents in urban areas. Data from the Czech Republic have shown a sim- ilar trend in the last decade [4]. VRUs generally have a high degree of freedom of movement in an ur- ban environment and tend to break traffic rules, thus 53 https://doi.org/10.14311/APP.2024.51.0053 https://creativecommons.org/licenses/by/4.0/ https://www.cvut.cz/en O. Piksa, J. Ritter, P. Drbohlav, T. Petrán Acta Polytechnica CTU Proceedings showing a greater propensity for accidents and injuries. There are even differences in the behaviour in different cities [5–8]. The statistics show that achieving the Vision 0 goal by 2050 will be difficult. Despite huge progress in the field of integrated ve- hicle safety and autonomous driving, collision-free transit and interaction with VRUs in urban traffic remains a challenge [9]. In any case, vulnerable users often cannot avoid death or serious injuries in a col- lision, even at relatively low speeds, when accident events in urban areas occur. Recognizing VRUs is a much more difficult task for a vehicle, e.g. as part of the functions of the AEB system, than recognizing other vehicles. 2. Research Most cities are equipped with relatively large camera surveillance systems. Such a developed network offers the opportunity to use video analysis in many areas dealing with traffic safety. There are currently not a lot of research and data sets available in the Czech Republic, which may create a gap due to the insuffi- cient coverage of input data needs for the development of advanced driver assistance systems (ADAS), Auto- mated Vehicles (AVs) and cooperative environments. Video analysis in the form of trajectories and speed profiles provides a description of the behaviour of individual road users at the micro level, which is a key input for many applications, such as assessment of the safety and efficiency of transport systems and infrastructure, or calibration of behavioural models, etc. This is confirmed by the conclusions of the extensive InDeV Horizon 2020 study [10]. The outputs of the InDeV project, just like Ahmed et al. [11], describe errors in police accident records that occur when mea- suring the accident site, information on injury rates, vehicles, roadway conditions and the specific environ- ments. The described inaccuracies have a significant impact on any analysis. At the end, the authors rec- ommend greater use and development of a method for studying traffic conflicts (e.g. Swedish Traffic Con- flict Technique, DOCTOR – Dutch Objective Con- flict Technique for Operation and Research) using video analysis. Data obtained from vehicle on-board cameras and infrastructure as one of the 4 pillars of data sources [12–17] are reported by the consortium P.E.A.R.S. [18], which addresses the development of the ISO standard for traffic safety assessment and for active safety technologies integrated into vehicles using virtual simulation. The reason for obtaining additional input data for development, testing and homologation is the fact that ADAS systems do not cover 100% of all scenarios, while AV system require 100% coverage of ideally all possible real critical sce- narios [19]. The verification of the functions of ADAS systems vs. VRU in independent testing is described in PEDICRASH and AAA tests and studies, the re- Figure 1. Ideal video analysis system. sults of which show limited effectiveness on selected complex scenarios [20, 21]. Real critical scenarios, situations and scenes require high-quality digitization, with the interconnection of driving data, and together with the application of methods using surrogate measures of safety (SMoS), they enable quick proactive solutions to understand all the causal contexts of accident and near-accident events. The given environment of critical situations can be reconstructed through scanning using multi-image photogrammetry. The output is a cloud of points that is used to create 3D models. The importance in terms of accuracy and time savings of accident site measurements with its subsequent virtualization when using modern measurement techniques by the police and forensic science is indicated, for example, in the papers [22–24]. Other uses are reported, for example, in papers in the field of research on traffic flows [25] or road infrastructure safety [26, 27]. From our own experience, we can mention the application for in-depth research on the transport safety made by teams such as UFO (VW), VDB (Skoda Auto), AARU (Audi). Currently, some service departments of the Transport Police of the Czech Republic already use modern technology, such as UAV photogrammetry and ground laser scanning. Video analysis including kinematic data and trajec- tories was used, for example, in research within the PROSPECT project [28], a study from Vancouver, British Columbia, focused on cyclists [29], or an anal- ysis of pedestrian behaviour using walk parameters in Nanjing, China [30]. An innovative approach to data collection was ap- plied in the Ko-PER project containing data from 8 infrastructure cameras and 14 laser scanners. The data contain sequences with object markings, object classes and hundreds of road users [31]. The ideal video analysis processing workflow is described in Fig- ure 1. 54 vol. 51/2024 Digital reconstruction of real (near) accidents of VRU . . . Figure 2. Comparison of tyre marks from distorted and undistorted images. Figure 3. Scheme of a synthetic calibration objects. The detection of road users is dependent on the complexity of the road traffic environment and is not easy to handle, so it is in the interest of many stud- ies [32–39]. The existence of publicly available data sets and algorithms is important for the development and testing of detectors [40–44]. A very important part of image analysis is rectifica- tion; paper [45], for example, evaluates measurement errors based on lens distortion and its impact on pho- togrammetric measurement of the distorted image; Figure 2 shows how the difference in tyre track mea- surement can cause a difference in the final results of the accident analysis. Another study [46] addresses the measurement ac- curacy affected by calibration with respect to the ratio of area of the calibration object at the accident site or the location of the calibration object using a cluster of small flexible objects. Figure 3 shows a scheme of calibration of the object from the calculation of the homographic matrix H and Figure 4 shows a rectified track image. Another paper evaluates accuracy using the PC Rect software, see Figure 5 [47]. Trajectory extraction – object tracking with their detection has several approaches [48]. • Manual methods – the analyst will mark each partic- ipant based on the specified frequency. The reason for use compared to the automated method is high accuracy, at the same time it does not address the is- sue of possible occlusion [49], incorrect classification, etc. [50]. However, the extraction may be subject Figure 4. Rectified image of traffic accident. Figure 5. PC Rect software. to human error and the number of observations is limited. • Automated methods – they process the image to identify and track moving objects in sequential video frames. Compared to the manual method, auto- mated methods are not time consuming and there is no human error. It is possible to process a larger volume of trajectories. The problem is the aforemen- tioned occlusion, changes in light, the movement of shadows or the movement of larger groups. Many current studies deal with the monitoring and prediction of VRU trajectories [50–53]. The basic approaches are shown in Figure 6. The above procedures are applied to a wide range of software tools developed for trajectory extraction. [53– 60]. Most of the similar tools do not explicitly provide for use by other analysts. Speed measurement There are many different approaches to calculating speed with the application of different methods according to the purpose of the research [61]. For example, the method applied in the Prospect project ensured results among the derived trajectories with a relative accuracy of 0.5–1m [62]. Another study [63] at low speeds applies the cross ratio method from projective geometry with σ 1.3% and ∆ 1.5%. Another article in the field of forensic science pro- poses a method of estimating the speed of a vehicle in 55 O. Piksa, J. Ritter, P. Drbohlav, T. Petrán Acta Polytechnica CTU Proceedings Figure 6. Basic approaches to VRU trajectory pre- diction. Figure 7. T-Analyst. a video using a virtual plane and a virtual reference line (VSEM) [64]. The following article suggests re- constructing accidents from CCTV recordings using volumetric kinetic mapping (VKM) [65]. For an overview, we present freely available software applications available from websites with the function of trajectory extraction from video recordings [66]. T-Analyst This is a database solution that makes it possible to select conflicts according to classification. It contains tools for placing predefined shapes (models) in the image, to extract trajectories or derive the speed of the monitored subjects/objects at a certain time (see in Figure 7). RUBA (Road User Behaviour Analysis) The basis of the approach is a set of detectors that can be connected by logical rules. They are activated by movement and they highlight conflict situations in red. Under favourable conditions, it is possible to remove up to 90% of the original footage that does not contain relevant information (see in Figure 8). Traffic Intelligence project It provides a set of tools for the detection, tracking and classification of road users using a tracking algorithm mapping trajectory data and individual interactions (see in Figure 9). Figure 8. RUBA. Figure 9. Traffic Intelligence project. Figure 10. STRUDL. STRUDL: Surveillance Tracking Using Deep Learning This approach is used as part of the In- DeV project under Horizon 2020 as an open-source and free framework for videos captured by stationary cameras. It uses an object detector using deep learn- ing to track and trace movement in the scene (see in Figure 10). SafeCross It describes an open-source framework for proactive risk assessment of individual accidents. Cyclist detection and tracking is done automatically using advanced computer vision techniques and deep learning (see in Figure 11). Last but not least, the use of virtual reality is be- coming increasingly important for examining critical scenarios. Especially in behavioural research involving HMI. For example, article [67] represents empirical re- search evaluating different HMI concepts of behaviour 56 vol. 51/2024 Digital reconstruction of real (near) accidents of VRU . . . Figure 11. SafeCross. Figure 12. Interactions in situations with multiple pedestrians. Figure 13. The test process. for two pedestrians, see Figure 12. Or study [68] deal- ing with decision-making processes when crossing the road in VRU vs. inbound vehicle interactions. The process of the test can be seen in Figure 13. 3. Method Our way of looking at road safety in terms of pre- venting road accidents is unusual, and thus brings a unique point of view on the whole issue. The cre- ated proof of concept called virtual accident experi- ence for real survival uses digital reconstruction of real accident event(s) all the way to the final use of HMD technology. In addition to the methodological procedure, we have created a software application written in Python, designed specifically for the Czech Police. The reason for our own In-house solution is the unavailability of SW licences for programmes to be commonly used by the units of the integrated rescue system. We attach a comparison of the conventional Virtual CRASH method with our method at the end of the chapter; at the same time it serves a validation function. By attaching a roadmap of reconstructed real (near-)accidents, we demonstrate the application in practice. Specifically, this covers the following loca- tions in the Czech Republic: Prague, Liberec, Česká Lípa, and Jaroměř. 3.1. Methodological approach The methodology for evaluating the 2D movement of an object is based on the image analysis of a recording from a single camera. This is a simple analysis, where the first prerequisite is the movement of the object on the horizontal XY plane. The effect of the Z-axis (bottom-up direction) is not taken into account. The advantage of the whole approach is its use even if we only know the top view of the situation, the bird’s eye view/aerial view. For evaluation purposes, only an image from publicly available sources of various map materials, e.g. google.com/maps or mapy.cz, is sufficient. However, these data are burdened by distortion caused by their acquisition at angles of up to 30°. Therefore, for the subsequent refinement of the results, we supplement the map data set with laser scanning of the area; this step will give us the benefits of a real orthographic map. The accuracy of the measurement reaches the maximum uncertainty in units of centimetres, typically up to 2 cm on an area of 600m2. The quality of the orthographic map can be exported in high resolution. The basic information for evaluation also includes the sampling frequency of the camera or video recording (FPS). The quality of the images is also influenced by the resolution of the camera, the lens used and the lighting conditions. The resolution of the orthophoto map, in other words the size of the scale – the ratio of the number of pixels to distances [pix.m−1], can also be considered a decisive influence. When using scanner measurement, we can achieve hundreds to thousands of pixels per metre. The methodology is usable for both static and dy- namic (on-board) camera recordings. The principle of evaluation is based on the search for optical pairs of points (places) that are identical for the video record- ing and the orthophoto map. These points introduce a dimension into the image (there must be a sufficient number of them to achieve better results), which is defined by the orthophoto map. It is advisable to have the distribution of control points in the image especially in those places where the objects of inter- est move. The control points must be positioned on the XY plane to maintain clarity between the 2D orthomap and the image. Specifically, these are un- ambiguous contrasting elements, such as corners at pedestrian crossings, footings of traffic signs, curb joints, etc. Since the influence of perspective is not linear, it is advisable to place control points in both directions of the image. The impact of perspective on the results is then significant. We achieve the best results in the case of a record taken from a bird’s eye view, similar to an orthographic one. On the contrary, 57 O. Piksa, J. Ritter, P. Drbohlav, T. Petrán Acta Polytechnica CTU Proceedings in the case of shooting from a small (low) angle, the quality of the results is reduced; this is the case of an on-board camera, where the angle between the pad and the object is relatively small. The next step is to monitor the movement of the object itself in the image – object tracking. Due to the aforementioned limitation of movement on the XY plane, it is neces- sary to mark the tracked object at a place as close as possible to this plane, e.g., the point of contact of the tyre with the road. Tracking is performed manu- ally on selected key frames (for reasons of accuracy); the timeline is defined by the sampling rate (FPS). The 2D coordinates obtained in this way in the video recording are converted into their equivalents in the 2D orthophoto map using linear interpolation. The obtained 2D motion coordinates are interlaced with a second-order polynomial across three points, due to smooth transitions between individual key frames. This gives us a two-dimensional description of the movement of objects in space – their position, speed and direction. The on-board record is evaluated with the difference that the above procedure is applied to each frame, or rather a suitable frame, in which control points can be defined with certainty. This eliminates the need to determine the position of the object, because this is the position of the camera itself. The results are influenced by the optical parameters of the camera itself (focal length of the lens and frame size = FOV). 3.2. Created software application When developing the software tool, we referred to the methodology described in Section 3.1, further dividing it into 3 consecutive phases, shown in Figures 14, 15 and 16. For a clear illustration, we attach a practi- cal reconstruction of one of the locations – Jaroměř (shown in Figures 17– 23. This deals with the measurement of an accident site – ortho plan, from the accident site survey using more ground or aerial photogrammetry. The aim is to convert the marked tracks into 2D coordinates in scale using manual input and automatic detection based on the set parameters in the image. The format of the text output is compatible with the processes and technical equipment of the police used for subsequent completion of the report on the traffic accident. The application was optimized and subsequently tested. Implementation of an application for measuring track coordinates in a 2D orthophoto plan using au- tomatic detection has brought the following function- alities (the output is an executable version of the application that has been tested): • Import of the requested ortho plan, • Manual entry of track position, • Automatic track detector using a defined search area, including parameter settings for detection (colour, threshold, gradient), • Scale definition, • Definition of the initial point of measurement, • Orientation of the coordinate system using two points, • Option to edit the description and annotation of each defined track, • Option to group multiple defined points into one track, • Option to introduce GPS 2D coordinates using a TXT file, • Export of coordinates into a text file and a 2D ortho plan of the scene. It includes the implementation of an application for dynamic analysis of the course of a traffic accident. The tool makes it possible to use both stationary and on-board camera recordings. The result is a visual interpretation of the trajectory of the traffic accident participants, including information on their speeds. The application and the implementation of annota- tions were tested again. At the same time, a methodol- ogy for defining primitives for the automatic creation of virtual reality was proposed and approved. In addition, the ortho plan of the traffic accident scene was supplemented with the trajectories and speeds of the traffic accident participants. This allows us to bring important information and a completely new or different view of the specific traffic accident. It also provides data for the “live” virtual reality scene. As part of the implementation of an application for dynamic analysis of a traffic accident using a single camera, the following objectives have been achieved: • Import of AVI, MP4 or MOV video, • Basic video editing – selection of the area of interest (video clipping), • Definition of pairs of points designed to evaluate the position in space using known coordinates from the ortho plan, • Definition of objects and their points, • Tracking of points in time, • Evaluation of 2D coordinates of points and objects depending on time, • Export of coordinates into a text file, • Export into a 2D ortho plan of the traffic accident scene. Export intended for visualizations in a virtual scene. It includes the development and implementation of an application for automatic creation of a 3D scene of the traffic accident scene based on an orthophoto map supplemented with primitives. The application and the implementation of annotations were tested. Creation of methodological materials for all 3 phases. Application for measuring track coordinates in a 2D orthophoto plan using automatic detection and appli- cation of video analysis of the accident event record. An application has been implemented for automatic 58 vol. 51/2024 Digital reconstruction of real (near) accidents of VRU . . . Figure 14. Phase 1 – Measuring tracks Figure 15. Phase 2 – Dynamic analysis of movement 59 O. Piksa, J. Ritter, P. Drbohlav, T. Petrán Acta Polytechnica CTU Proceedings Figure 16. Phase 3 – Creating virtual reality (VR) 60 vol. 51/2024 Digital reconstruction of real (near) accidents of VRU . . . Figure 17. Defining scale. Figure 18. Defining the initial point of the measurement (IPM) and the direction of the coordinate system. Figure 19. Pairing common points. 61 O. Piksa, J. Ritter, P. Drbohlav, T. Petrán Acta Polytechnica CTU Proceedings Figure 20. Results of dynamic video analysis. Figure 21. Creating a 3D virtual scene. Figure 22. Dynamic VR scene object library. 62 vol. 51/2024 Digital reconstruction of real (near) accidents of VRU . . . Figure 23. Vehicle view in the VR scene. creation of a 3D scene of the traffic accident scene based on an orthophoto map supplemented with prim- itives. Anonymized video outputs from virtual reality. As part of the implementation of the application, the planned objectives have been achieved: • Retrieval of the entered objects from the XML source and additional information, • Setting of the loaded objects to the specified posi- tion and the required rotation, • Editing of the basic shape of the specified objects based on the control curves, • Setting of the tracks and speeds of moving objects, • Loading of specified tracking points, both in space and tied to moving objects, • Setting of the scene background and lighting, • Creation of control and information systems and means for previewing the scene. Example results can be seen in Figures 24 and 25, which depict a reconstructed situation from Liberec, Malé náměstí. Subsequently, validation of the ob- tained results was carried out on another dataset from Praha, Smetanovo nábřeží, as illustrated in Fig- ures 26, 27 and 28. 3.3. Validation of results from Praha, Smetanovo nábřeží 4. Discussion The quality of the results is influenced by a number of factors, which are: camera/recording resolution, lens distortion and its type, angle of view of the scene, direction and size of the object’s movement relative to the rotation of the frame (effect of perspective), record- ing sampling rate (FPS) and its stability, contrast between the monitored object and the environment, number of control points, quality of the orthophoto map (number of pix.m−1), video codec used. Each of these factors will more or less influence the result. From the point of view of a fundamental improve- ment of the results of the methodology it would be advisable to: • To perform a lens calibration to compensate for image distortion, • More effective placement of control points through- out the image. Ideally a regular grid with a size of e.g. 2 × 2m, • Digitize the entire area of interest in order to obtain 3D data. Then define control points directly in the 3D space, • Use the Z-axis information (3D data) to evaluate the position of the object, • Try to avoid obscuring significant points in the image (camera position relative to static objects in the camera frame), • Interleave tracked points using more complex al- gorithms – 3rd order polynomial using 5 or more points, etc. 5. Conclussion We have achieved the planned results of the devel- opment of a software tool and methodology for the processing of marked tracks from the scene of a traffic accident, taken with multi-image ground or air pho- togrammetry, 3D scanning, or other methods whose output is an ortho plan (rectified photographic image). The SW application makes it possible to create scenes for virtual reality based on data sets from real traffic accidents. Our effort has also been to identify factors affecting the quality of output data for the needs of safety and efficiency of transport systems, infrastruc- ture or calibration of behavioural models (HMI). At the same time, nationwide training of representatives of the Transport Police and the Czech Police Head- quarters has taken place in 14 regions in the use of the software tool and methodology for processing the location of a traffic accident, as well as nationwide training of representatives of the Communication and External Relations Department of the Czech Police Headquarters in the use of the software application and methodology for automatic creation of a 3D scene of the traffic accident based on an orthophoto map supplemented with primitives. In cooperation with the Czech Police and the Vision 0 platform, preven- tive actions using VR simulations based on real traffic accidents are still ongoing. Finally, the paper has presented the verification of a method to obtain parameters such as speed and dis- tance at a certain time, thus enabling the acquisition of SMO identifiers even when using a lower-quality video recording. 6. Future work A follow-up study of traffic conflicts will be carried out using appropriate camera systems. The examined 63 O. Piksa, J. Ritter, P. Drbohlav, T. Petrán Acta Polytechnica CTU Proceedings Figure 24. Dynamic motion analysis results. Figure 25. Digital reconstruction (left) and simulation in Unity and VR (right). Figure 26. Accident analysis – video record. 64 vol. 51/2024 Digital reconstruction of real (near) accidents of VRU . . . Figure 27. Accident analysis – conventional method. Figure 28. Accident analysis – methods comparison. critical situations will include primarily collisions, po- tential collisions and traffic conflicts between pedes- trians and vehicles, secondarily between pedestrians and cyclists, cyclists and vehicles, e-scooter riders, e- micromobility users, and other road users. A holistic approach will be applied to the acquired video data set to detect VRU intentions using cooperative meth- ods. Intention detection will consist of basic primitive motion prediction, e.g. standing, moving or rotating. The subsequent trajectory estimate will be evaluated and categorized. This will be preceded by the selec- tion of a suitable detection algorithm based on the DL architecture or a combination using the Bayesian BDL probability with the application of e.g. Fast-RCNN, R-FCN, Faster R-CNN, YOLOv5 or SSD. Most accidents are due to human error. Therefore, it is advisable to add in the future a system view of the communication interface between the operator (driver) and the artificial system (vehicle, its parts). This will be achieved by using knowledge from the Driving Simulations Research Group at CTU, devel- oping interactive vehicle simulators, scenario genera- tion technologies and experiments using appropriate methods. Instruments working with psychophysical measures, advanced analytical and classification tools can add extra value to the results [69–78]. Acknowledgements This article has been supported by the project “Virtual Survival of an Accident for Real Survival”, financed by the Damage Prevention Fund of the Czech Insurers’ Bureau. The successful extension of the project has been then covered from Aurel CZ’s own resources. References [1] K. Gschwendtner, P. Feig, M. Kiss, M. Lienkamp. Prospective estimation of the effectiveness of driver assistance systems in property damage accidents. In 24th ESV 2015 International Technical Conference on the Enhanced Safety of the Vehicles. 2015. Paper Number 15-0037. [2] World Health Organization. Global status report on road safety, 2018. 65 O. Piksa, J. Ritter, P. Drbohlav, T. Petrán Acta Polytechnica CTU Proceedings [3] Directorate-General for Mobility and Transport. Road safety in the EU: fatalities below pre-pandemic levels but progress remains too slow, 2018. [4] Centre for Transport Research, v. v. i. (CDV). Bezpečnost silničního provozu v EU: počet úmrtí v roce 2021 zůstává výrazně pod úrovní před pandemií [In Czech; Road safety in the EU: number of deaths in the year 2021 remains well below pre-pandemic levels], 2022. [5] L. Hell, J. Sprenger, M. Klusch, et al. Pedestrian behavior in Japan and Germany: A review. In 2021 IEEE Intelligent Vehicles Symposium (IV), pp. 1529–1536. 2021. https://doi.org/10.1109/IV48863.2021.9575809 [6] G. Solmazer, D. Azık, G. Fındık, et al. Cross-cultural differences in pedestrian behaviors in relation to values: A comparison of five countries. Accident Analysis & Prevention 138:105459, 2020. https://doi.org/10.1016/j.aap.2020.105459 [7] P. Olszewski, P. Szagała, D. Rabczenko, A. Zielińska. Investigating safety of vulnerable road users in selected EU countries. Journal of Safety Research 68:49–57, 2019. https://doi.org/10.1016/j.jsr.2018.12.001 [8] X. Jiang, W. Wang, K. Bengler, W. Guo. Analyses of pedestrian behavior on mid-block unsignalized crosswalk comparing Chinese and German cases. Advances in Mechanical Engineering 7(11):1687814015610468, 2015. https://doi.org/10.1177/1687814015610468 [9] SAE. Automated driving – levels of driving automation are defined in new sae international standard J3016. Tech. rep., 2016. [10] InDeV Horizon 2020. In-depth understanding of accident causation for vulnerable road users, 2016. Project No. 635895. [11] A. Ahmed, A. F. M. Sadullah, A. S. Yahya. Errors in accident data, its types, causes and methods of rectification – analysis of the literature. Accident Analysis & Prevention 130:3–21, 2019. https://doi.org/10.1016/j.aap.2017.07.018 [12] F. Guo, S. G. Klauer, M. T. McGill, T. A. Dingus. Evaluating the relationship between near-crashes and crashes: Can near-crashes serve as a surrogate safety metric for crashes?, 2010. [2023-11-21]. https://www.nhtsa.gov/sites/nhtsa.gov/files/ documents/dot_hs_811_382.pdf [13] GIDAS. [2023-11-21]. Https://Www.Gidas.Org/Start-En.Html [14] NHTSA. National Automotive Sampling System. [2023-11-21]. Https://Www.Nhtsa.Gov/Crash-Data- Systems/National-Automotive-Sampling-System [15] European Commission. CARE database. [2023-11-21]. Https://Road-Safety.Transport.Ec. Europa.Eu/Statistics-and-Analysis/Methodology- and-Research/Care-Database_en [16] ITARDA. [2023-11-21]. Https://Www.Itarda.or.Jp/English/ [17] P. Lesire, F. Léopold, C. Chauvel, et al. Implication of children in road accidents in France in 2011. In 24th International Technical Conference on the Enhanced Safety of Vehicles (ESV). 2015. [18] P. Wimmer, H. Weber, H. Chajmowicz, et al. Harmonized approaches for baseline creation in prospective safety, performance assessment of driving automation systems. Paper Number 23-0032. [19] Cetran scenario categories. [2023-11-21]. https://cetran.sg/wpcontent/uploads/2020/01/ REP200121_Scenario_Categories_v1.7.pdf [20] M. Ucińska, M. Pełka. The effectiveness of the AEB system in the context of the safety of vulnerable road users. Open Engineering 11(1):977–993, 2021. https://doi.org/10.1515/eng-2021-0097 [21] AAA, Inc. AAA NewsRoom, 2019. [2023-11-21]. https://newsroom.aaa.com/asset/pedestrian- detection-full-research-report-automotive-oct- 2019/ [22] R. Kamnik, M. Nekrep Perc, D. Topolšek. Using the scanners and drone for comparison of point cloud accuracy at traffic accident analysis. Accident Analysis & Prevention 135:105391, 2020. https://doi.org/10.1016/j.aap.2019.105391 [23] J. Wang, Z. Li, Fan ying, et al. Reconstruction of a real-world car-to-pedestrian collision using geomatics techniques and numerical simulations. Journal of Forensic and Legal Medicine 91:102433, 2022. https://doi.org/10.1016/j.jflm.2022.102433 [24] I. Stáňa, S. Tokař, K. Bucsuházy, M. Bilík. Comparison of utilization of conventional and advanced methods for traffic accidents scene documentation in the Czech Republic. Procedia Engineering 187:471–476, 2017. https://doi.org/10.1016/j.proeng.2017.04.402 [25] P. Garcia-Aunon, J. J. Roldán, A. Barrientos. Monitoring traffic in future cities with aerial swarms: Developing and optimizing a behavior-based surveillance algorithm. Cognitive Systems Research 54:273–286, 2019. https://doi.org/10.1016/j.cogsys.2018.10.031 [26] L. Inzerillo, G. Di Mino, R. Roberts. Image-based 3D reconstruction using traditional and UAV datasets for analysis of road pavement distress. Automation in Construction 96:457–469, 2018. https://doi.org/10.1016/j.autcon.2018.10.010 [27] C. Zhang. An UAV-based photogrammetric mapping system for road condition assessment. In The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, p. 627–632. 2008. [28] M.-P. Bruyas, A. Aparicio, M. Kosel, et al. Naturalistic observations to investigate conflicts between drivers and VRUs in the PROSPECT project. In SV 2017 – 25th International Technical Conference on the Enhanced Safety of Vehicles, pp. 1–11. 2017. [29] F. M. Puscar, T. Sayed, A. Y. Bigazzi, M. H. Zaki. Multimodal safety assessment of an urban intersection by video analysis of bicycle, pedestrian, and motor vehicle traffic conflicts and violations. In 97th Annual Meeting of the Transportation Research Board, pp. 1–14. 2017. [30] Y. Guo, T. Sayed, M. H. Zaki. Automated analysis of pedestrian walking behaviour at a signalised intersection in China. IET Intelligent Transport Systems 11(1):28–36, 2017. https://doi.org/10.1049/iet-its.2016.0090 66 https://doi.org/10.1109/IV48863.2021.9575809 https://doi.org/10.1016/j.aap.2020.105459 https://doi.org/10.1016/j.jsr.2018.12.001 https://doi.org/10.1177/1687814015610468 https://doi.org/10.1016/j.aap.2017.07.018 https://www.nhtsa.gov/sites/nhtsa.gov/files/documents/dot_hs_811_382.pdf https://www.nhtsa.gov/sites/nhtsa.gov/files/documents/dot_hs_811_382.pdf Https://Www.Gidas.Org/Start-En.Html Https://Www.Nhtsa.Gov/Crash-Data-Systems/National-Automotive-Sampling-System Https://Www.Nhtsa.Gov/Crash-Data-Systems/National-Automotive-Sampling-System Https://Road-Safety.Transport.Ec.Europa.Eu/Statistics-and-Analysis/Methodology-and-Research/Care-Database_en Https://Road-Safety.Transport.Ec.Europa.Eu/Statistics-and-Analysis/Methodology-and-Research/Care-Database_en Https://Road-Safety.Transport.Ec.Europa.Eu/Statistics-and-Analysis/Methodology-and-Research/Care-Database_en Https://Www.Itarda.or.Jp/English/ https://cetran.sg/wpcontent/uploads/2020/01/REP200121_Scenario_Categories_v1.7.pdf https://cetran.sg/wpcontent/uploads/2020/01/REP200121_Scenario_Categories_v1.7.pdf https://doi.org/10.1515/eng-2021-0097 https://newsroom.aaa.com/asset/pedestrian-detection-full-research-report-automotive-oct-2019/ https://newsroom.aaa.com/asset/pedestrian-detection-full-research-report-automotive-oct-2019/ https://newsroom.aaa.com/asset/pedestrian-detection-full-research-report-automotive-oct-2019/ https://doi.org/10.1016/j.aap.2019.105391 https://doi.org/10.1016/j.jflm.2022.102433 https://doi.org/10.1016/j.proeng.2017.04.402 https://doi.org/10.1016/j.cogsys.2018.10.031 https://doi.org/10.1016/j.autcon.2018.10.010 https://doi.org/10.1049/iet-its.2016.0090 vol. 51/2024 Digital reconstruction of real (near) accidents of VRU . . . [31] E. Strigel, D. Meissner, F. Seeliger, et al. The Ko-PER intersection laserscanner and video dataset. In 17th International IEEE Conference on Intelligent Transportation Systems (ITSC), pp. 1900–1901. 2014. https://doi.org/10.1109/ITSC.2014.6957976 [32] B. Leibe, K. Schindler, N. Cornelis, L. Van Gool. Coupled object detection and tracking from static cameras and moving vehicles. IEEE Transactions on Pattern Analysis and Machine Intelligence 30(10):1683–1698, 2008. https://doi.org/10.1109/TPAMI.2008.170 [33] T. Schoenemann, D. Cremers. Globally optimal shape- based tracking in real-time. In 2008 IEEE Conference on Computer Vision and Pattern Recognition, pp. 1–6. 2008. https://doi.org/10.1109/CVPR.2008.4587443 [34] T. Zhao, R. Nevatia. Tracking multiple humans in complex situations. IEEE Transactions on Pattern Analysis and Machine Intelligence 26(9):1208–1221, 2004. https://doi.org/10.1109/TPAMI.2004.73 [35] P. Mannion. Vulnerable road user detection: state-of-the-art and open challenges, 2019. https://doi.org/10.48550/arXiv.1902.03601 [36] R. Benenson, M. Omran, J. Hosang, B. Schiele. Ten years of pedestrian detection, what have we learned? In L. Agapito, M. M. Bronstein, C. Rother (eds.), Computer Vision – ECCV 2014 Workshops, pp. 613–627. Springer International Publishing, Cham, 2015. https://doi.org/10.1007/978-3-319-16181-5_47 [37] S. Zhang, R. Benenson, M. Omran, et al. How far are we from solving pedestrian detection? In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 1259–1267. 2016. https://doi.org/10.1109/CVPR.2016.141 [38] P. Dollar, C. Wojek, B. Schiele, P. Perona. Pedestrian detection: An evaluation of the state of the art. IEEE Transactions on Pattern Analysis and Machine Intelligence 34(4):743–761, 2012. https://doi.org/10.1109/TPAMI.2011.155 [39] M. He, H. Luo, Z. Chang, B. Hui. Pedestrian detection with semantic regions of interest. Sensors 17(11):2699, 2017. https://doi.org/10.3390/s17112699 [40] S. Zernetsch, H. Reichert, V. Kress, et al. Cyclist intention detection: A probabilistic approach, 2021. https://doi.org/10.48550/arXiv.2104.09176 [41] M. Braun, S. Krebs, F. Flohr, D. M. Gavrila. The EuroCity persons dataset: A novel benchmark for object detection, 2018. https://doi.org/10.48550/arXiv.1805.07193 [42] S. Zhang, R. Benenson, B. Schiele. CityPersons: A diverse dataset for pedestrian detection. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 4457–4465. 2017. https://doi.org/10.1109/CVPR.2017.474 [43] A. Geiger, P. Lenz, R. Urtasun. Are we ready for autonomous driving? The KITTI vision benchmark suite. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pp. 3354–3361. 2012. https://doi.org/10.1109/CVPR.2012.6248074 [44] X. Li, F. Flohr, Y. Yang, et al. A new benchmark for vision-based cyclist detection. In 2016 IEEE Intelligent Vehicles Symposium (IV), pp. 1028–1033. 2016. https://doi.org/10.1109/IVS.2016.7535515 [45] W. Neale, D. Hessel, T. Terpstra. Photogrammetric measurement error associated with lens distortion. Tech. rep., 2011. SAE Technical Paper 2011-01-0286. https://doi.org/10.4271/2011-01-0286 [46] Q. Chen, H. Xu, L. Tan. Application of composite small calibration objects in traffic accident scene photogrammetry. PLoS ONE 10(5):e0127018, 2015. https://doi.org/10.1371/journal.pone.0127018 [47] P. Kohút, L. Macurová, M. Rédl, M. Ballay. Application of rectification method for processing of documentation from the place of road accident. The Archives of Automotive Engineering – Archiwum Motoryzacji 88(2):37–46, 2020. https://doi.org/10.14669/AM.VOL88.ART3 [48] H. Twaddle, T. Schendzielorz, O. Fakler, S. Amini. Use of automated video analysis for the evaluation of bicycle movement and interaction. In R. P. Loce, E. Saber, N. Lecky (eds.), Video Surveillance and Transportation Imaging Applications 2014, vol. 9026, p. 90260U. International Society for Optics and Photonics, 2014. https://doi.org/10.1117/12.2042487 [49] J. Pan, B. Hu. Robust occlusion handling in object tracking. In 2007 IEEE Conference on Computer Vision and Pattern Recognition, pp. 1–8. 2007. https://doi.org/10.1109/CVPR.2007.383453 [50] J. Gwak. Multi-object tracking through learning relational appearance features and motion patterns. Computer Vision and Image Understanding 162:103–115, 2017. https://doi.org/10.1016/j.cviu.2017.05.010 [51] C. Zhang, C. Berger, M. Dozza. Social-IWSTCNN: A Social Interaction-Weighted Spatio- Temporal Convolutional Neural Network for pedestrian trajectory prediction in urban traffic scenarios. In 2021 IEEE Intelligent Vehicles Symposium (IV), pp. 1515–1522. 2021. https://doi.org/10.1109/IV48863.2021.9575958 [52] S. Zernetsch, O. Trupp, V. Kress, et al. Cyclist trajectory forecasts by incorporation of multi-view video information. In 2021 IEEE International Smart Cities Conference (ISC2), pp. 1–7. 2021. https://doi.org/10.1109/ISC253183.2021.9562857 [53] M. M. Karim, D. Doell, R. Lingard, et al. A region-based deep learning algorithm for detecting and tracking objects in manufacturing plants. Procedia Manufacturing 39:168–177, 2019. https://doi.org/10.1016/j.promfg.2020.01.289 [54] Iteris. Pedestrian & cyclist safety. [2023-11-21]. https://www.iteris.com/oursolutions/pedestrian- cyclist-safety [55] Viscando. Traffic. [2023-11-21]. https://viscando.com/applications/traffic/ [56] Data from sky. [2023-11-21]. https://datafromsky.com/ [57] COWI. [2023-11-21]. https://www.cowi.com/ solutions/planning/accident-analysis-denmark 67 https://doi.org/10.1109/ITSC.2014.6957976 https://doi.org/10.1109/TPAMI.2008.170 https://doi.org/10.1109/CVPR.2008.4587443 https://doi.org/10.1109/TPAMI.2004.73 https://doi.org/10.48550/arXiv.1902.03601 https://doi.org/10.1007/978-3-319-16181-5_47 https://doi.org/10.1109/CVPR.2016.141 https://doi.org/10.1109/TPAMI.2011.155 https://doi.org/10.3390/s17112699 https://doi.org/10.48550/arXiv.2104.09176 https://doi.org/10.48550/arXiv.1805.07193 https://doi.org/10.1109/CVPR.2017.474 https://doi.org/10.1109/CVPR.2012.6248074 https://doi.org/10.1109/IVS.2016.7535515 https://doi.org/10.4271/2011-01-0286 https://doi.org/10.1371/journal.pone.0127018 https://doi.org/10.14669/AM.VOL88.ART3 https://doi.org/10.1117/12.2042487 https://doi.org/10.1109/CVPR.2007.383453 https://doi.org/10.1016/j.cviu.2017.05.010 https://doi.org/10.1109/IV48863.2021.9575958 https://doi.org/10.1109/ISC253183.2021.9562857 https://doi.org/10.1016/j.promfg.2020.01.289 https://www.iteris.com/oursolutions/pedestrian-cyclist-safety https://www.iteris.com/oursolutions/pedestrian-cyclist-safety https://viscando.com/applications/traffic/ https://datafromsky.com/ https://www.cowi.com/solutions/planning/accident-analysis-denmark https://www.cowi.com/solutions/planning/accident-analysis-denmark O. Piksa, J. Ritter, P. Drbohlav, T. Petrán Acta Polytechnica CTU Proceedings [58] GoodVision. Traffic data collection from video. [2023-11-21]. https://goodvisionlive.com/solutions/traffic- data-collection/ [59] intuVision. intuVision VA Traffic. [2023-11-21]. https://www.intuvisiontech.com/intuvisionVA_ solutions/intuvisionVA_traffic [60] Transoft Solutions. TrafxSAFE Brochure. [2023-11-21]. https://safety.transoftsolutions.com/trafxsafe/ [61] A. Laureshyn, H. Ardö, A. Svensson, T. Jonsson. Application of automated video analysis for behavioural studies: concept and experience. IET Intelligent Transport Systems 3:345–357, 2009. https://doi.org/10.1049/iet-its.2008.0077 [62] Z. Koppanyi, C. Toth, T. Soltesz. Deriving pedestrian positions from uncalibrated videos. In IGTF 2017 – Imaging & Geospatial Technology Forum 2017. 2017. [63] T. W. Wong, C. H. Tao, Y. K. Cheng, et al. Application of cross-ratio in traffic accident reconstruction. Forensic Science International 235:19–23, 2014. https://doi.org/10.1016/j.forsciint.2013.11.012 [64] J.-H. Kim, W.-T. Oh, J.-H. Choi, J.-C. Park. Reliability verification of vehicle speed estimate method in forensic videos. Forensic Science International 287:195–206, 2018. https://doi.org/10.1016/j.forsciint.2018.04.002 [65] E. Kolla, J. Ondruš, P. Vertal. Reconstruction of traffic situations from digital video-recording using method of volumetric kinetic mapping. The Archives of Automotive Engineering – Archiwum Motoryzacji 84(2):147–170, 2019. https://doi.org/10.14669/AM.VOL84.ART11 [66] ICTCT. Surrogate measures of safety. [2023-11-21]. https://www.ictct.net/smos/software/ [67] D. Dey, A. van Vastenhoven, R. H. Cuijpers, et al. Towards scalable eHMIs: Designing for AV-VRU communication beyond one pedestrian. In 13th International Conference on Automotive User Interfaces and Interactive Vehicular Applications, p. 274–286. 2021. https://doi.org/10.1145/3409118.3475129 [68] F. Soares, E. Silva, F. Pereira, et al. To cross or not to cross: Impact of visual and auditory cues on pedestrians’ crossing decision-making. Transportation Research Part F: Traffic Psychology and Behaviour 82:202–220, 2021. https://doi.org/10.1016/j.trf.2021.08.014 [69] P. Bouchner. Interactive driving simulators – history, design and their utilization in area of HMI research. International journal of systems applications, engineering & development 10:179–188, 2016. [70] M. Cenkner, J. Svoboda, P. Toman, P. Bouchner. Adaptation of the vehicle simulator on BEV and HEV, focused on the impact studies of new user interfaces on drivers. In 2023 Smart City Symposium Prague (SCSP), pp. 1–7. 2023. https://doi.org/10.1109/SCSP58044.2023.10146111 [71] D. Rojdestvenskiy, M. Cvetković, P. Bouchner. Real-time driver advisory system for improving energy economy based on advance driver assistant systems interface. In 2018 Smart City Symposium Prague (SCSP), pp. 1–6. 2018. https://doi.org/10.1109/SCSP.2018.8402655 [72] P. Bouchner, S. Novotný, D. Rozhdestvenskiy, et al. Driver’s aggressive behavior – experiments on the driving simulator. In Latest Trends in Energy, Enviromental and Development, pp. 148–153. 2014. ISBN 978-960-474-375-9. [73] P. Bouchner, M. Leso, A. Matlafus, et al. Additional in-car warning system for safer grade crossings. In Latest Trends in Energy, Enviromental and Development, pp. 127–131. 2014. ISBN 978-960-474-375-9. [74] P. Bouchner, S. Novotný. Secondary driving tasks experiments with use of eye-tracking device. In Road Safety and Simulation International Conference. 2013. [75] P. Bouchner, S. Novotný. Validation of passenger car mathematical model for training driving simulators. In Advances in Sensors, Signals, Visualization, Imaging and Simulation, pp. 269–273. 2012. ISBN 978-1-61804-119-7. [76] P. Bouchner, S. Novotný. Car dynamics model – design for interactive driving simulation use. In Recent Researches in Applied Informatics, pp. 285–289. 2011. ISBN 978-1-61804-034-3. [77] P. Bouchner, S. Novotný, R. Piekník, et al. Analysis of real car driving data for simulator development. In Proceedings of the International Conference TRANSTEC, pp. 254–261. 2007. ISBN 978-80-01-03782-9. [78] P. Bouchner. Simulator technology for research and training, 2011. Associate Professor Lecture. Czech Technical University in Prague. 68 https://goodvisionlive.com/solutions/traffic-data-collection/ https://goodvisionlive.com/solutions/traffic-data-collection/ https://www.intuvisiontech.com/intuvisionVA_solutions/intuvisionVA_traffic https://www.intuvisiontech.com/intuvisionVA_solutions/intuvisionVA_traffic https://safety.transoftsolutions.com/trafxsafe/ https://doi.org/10.1049/iet-its.2008.0077 https://doi.org/10.1016/j.forsciint.2013.11.012 https://doi.org/10.1016/j.forsciint.2018.04.002 https://doi.org/10.14669/AM.VOL84.ART11 https://www.ictct.net/smos/software/ https://doi.org/10.1145/3409118.3475129 https://doi.org/10.1016/j.trf.2021.08.014 https://doi.org/10.1109/SCSP58044.2023.10146111 https://doi.org/10.1109/SCSP.2018.8402655 Acta Polytechnica CTU Proceedings 51:53–68, 2024 1 Introduction 2 Research 3 Method 3.1 Methodological approach 3.2 Created software application 3.3 Validation of results from Praha, Smetanovo nábřeží 4 Discussion 5 Conclussion 6 Future work Acknowledgements References