Unsignalized crosswalks remain the most vulnerable scenario where pedestrians are exposed to the highest risks. With the imminent introduction of autonomous vehicles on public roads, safe encounters with pedestrians in these critical environments presents a significant challenge. Our study develops a rigorous methodology to quantitatively assess these dynamics in real-world mixed trafic conditions. We implemented a system that processes video data from on-street cameras to evaluate risks in vehicle-pedestrian interactions by computing key conflict measures, such as the Time-to-Collision (TTC). The analysis conducted at an unsignalized pedestrian crossing enabled a comparative evaluation between conventional and autonomous vehicles. Results highlight a higher incidence of severe conflicts in interactions with human-operated vehicles, suggesting that the cautious programming of autonomous vehicles can significantly contribute to pedestrian safety. Our findings also reveal an impact on the pedestrian decision-making process based on the type of vehicle approaching the crosswalk.
The integration of Autonomous Vehicles (AV) into public roads represents a significant advancement in
transportation technology. Although testing has been performed primarily in controlled environments
[
Despite the prevalence of crash-based metrics, non-crash events and conflict measures provide valuable safety insights [9]. Previous studies have employed micro-simulation techniques to study AV-pedestrian interactions, but these approaches have limitations in modeling behavioral aspects and environmental factors [10]. Real-environment observations of AV-pedestrian interactions remain scarce due to high costs and legal requirements. Recent technological advances in video recording and automated analysis have facilitated safety-related event studies [11, 12, 13, 14].
This paper represents an extended abstract of a recent published work [15]. Our study investigates interactions between pedestrians and autonomous shuttles with SAE level 4 in am urban environment with dense mixed trafic in an Italian city, providing comparative insights with HV-pedestrian interactions. We propose an integrated processing pipeline for automated conflict analysis from on-street camera recordings. This pipeline includes: (i) road user detection and tracking to compute kinematic information; (ii) data analysis to model vehicle-pedestrian interactions and detect potential conflicts; (iii) conflict measures computation (e.g., Time-To-Collision) and pedestrian decision-making analysis. Our methodology moves from real-world scenarios to a more consistent representation through a vehicle box-based 2D bird’s eye view that accounts for diferent vehicle sizes.
Trafic data was collected using low-cost action cameras (Garmin VIRB, 1080p HD, 30 fps) mounted on a 10.80 m telescopic pole near a hospital area characterized by mixed vehicle types. The cameras were positioned discreetly outside the roadway to minimize driver awareness and potential behavioural modifications. The collected footage was corrected for distortion via the MATLAB Camera Calibrator App to eliminate wide-angle lens errors, which would afect the correct positioning and detected speed of users in the scene. Object detection was performed using YOLOv5 [16] with COCO classes [17] (manually labelling the autonomous shuttle since it is not part of the available classes), while StrongSORT facilitated consistent object tracking across frames [18].
For measuring spatial interactions between pedestrians and vehicles, we converted from pixels to a geo-referenced X-Y coordinate system. As reported in Fig. 2, in this system, the X-axis runs parallel to the road, while the Y-axis runs parallel to pedestrian crossings. The origin point (0,0) is located at the bottom left corner of the zebra crossing. We tracked specific reference points of the bounding boxes: for pedestrians, the center point between their feet; for vehicles, the center of the front license plate. Position tracking occurred every 34 ms.
To enhance the accuracy of spatial and motion-related variables, kinetic data is denoised using the Simple Moving Average (SMA), which mitigates noise from camera oscillation and image discretization. The SMA computes the unweighted mean over a window of data points, given by: = 1
∑︁ =−+1 (1) where is the position, is the total number of data points, and is the width of the window. A sensitivity analysis led to adopting = 30 frames (1s), ensuring a balance between noise reduction and accuracy. Validation using AV onboard sensor data confirms the efectiveness of this approach. The estimated speed aligns closely with the recorded values, producing 2 = 0.95, MAE = 0.92 km/h, and RMSE = 1.4 km/h.
After acquiring the coordinates and kinematic data, including trajectory and speed at each timestamp, we isolated the segment where the pedestrian and the AV/HV were actively engaged in an interaction. We then moved from the original camera image (Fig. 2) to the proper 2D perspective of the scene (Section 2.2.1) to analyze TTC and pedestrian decision-making (Section 2.2.2 and Section 2.2.3, respectively).
The camera images and standard bounding boxes obtained with the object detection system are not adequate to simulate a 2D bird’s-eye view consistent with the theoretical formulation considered. For vehicle representation, a 2D-box is positioned using the center of the frontal plate as the reference point. We established a reference catalogue of standard vehicle dimensions (Table 1) to automatically assign each detected vehicle to its corresponding box size. We applied standardized dimensions based on the vehicle type to mitigate inaccuracies due to the camera’s perspective and measurement noise. The pedestrian is instead approximated as a single point in the 2D bird’s-eye view, corresponding to the bottom center of the bounding box, given the diference in scale relative to vehicles.
Interactions between road users are categorized into two types: undisturbed passages, where users maintain their course without altering speed or direction due to adequate spacing, and conflicts, which require at least one participant to take evasive action to avoid a collision. Conflicts can be classified as minor, where minor adjustments (such as a slight speed reduction or a shift of the trajectory) are suficient to avoid danger, or serious, where a collision is inevitable unless significant evasive maneuvers are made by one or more road users [19].
Time-To-Collision (TTC) is the time that separates two road users from a collision in the pre-event phase if the collision course and speed diference are maintained [ 20]. Given two general Road Users 1 at speed 1 and 2 at speed 2, they are on a collision course if there is at least a straight parallel to the vector of speed diference ∆ = 2 − 1 that from the target user 2 intersects 1. The trajectories of 1 and 2 generate a conflicting area: the expected place of the incident assuming that neither road user takes any evasive action. Since speed and position change over time, the Instantaneous Time-To-Collision (ITTC) is calculated by evaluating the TTC at each instant . The minimum ITTC value ( ) represents the severity of the conflict, with lower values indicating a higher risk [ 21, 22, 23]. Based on literature [24, 25, 26], we classify ≥ 3 s as no conflict, 1.5 ≤ < 3 s as slight conflicts, and < 1.5 s as serious conflicts. Although thresholds vary between studies, we adopt 1.5 s as the benchmark for serious conflicts, following established conventions [22, 27].
Building on the widely accepted definition of (I)TTC, this study employs a precise mathematical approach that considers, on a frame-by-frame basis, the spatial dimensions and orientation of the interacting road users, and the interaction region to determine whether the two users are on a potential collision path. Each road user (RU ) is represented in a two-dimensional space by: (i) 2D box-based model showing its bird’s-eye view; (ii) the RU reference point, defined by the (X, Y) coordinates of the center of the front head; (iii) two RU extreme points, represented by the highest and lowest y-coordinate values of the box.
In this work, we define the interaction region as the area that captures the influence of the target user’s spatial occupancy on the interaction dynamics. This region is used to assess whether the two road users are on a potential collision course and whether the computation of the ITTC is feasible.
Fig. 3 illustrates the scenario of two road users, 1 and 2, each represented by a box. When 2 is considered the target user, the interaction region is defined as the area bounded by two lines parallel to the relative velocity vector ∆ , each passing through the extreme points of 2. Whether 1 falls within this interaction region determines whether there is a collision course between the two road users.
As illustrated in Fig. 3a, at time 0, 1 lies within the interaction region of 2 and thus they are on a collision course. If no evasive actions are taken, a conflict is anticipated to occur in the conflicting area after a time given by Equation 2: =
||∆ || (2) where represents the current distance between the two road users, defined as the shortest segment parallel to ∆ (denoted ∆ ) connecting the potential collision points 1 and 2. We refer this time instant as the Interaction Time (IT).
In Fig. 3b, at time 1, the reduction in speed of 1 (1,1 < 1,0) leads to a reorientation and a change in the magnitude of ∆ (here ∆ 1). As a result, the interaction region shifts and 1 is not included anymore. In this scenario, there is no longer a collision course, and the ITTC approaches (a) (b) infinity, indicating no actual conflict. We refer to this specific time instant as the no-Interaction Time (no-IT).
From the collected images (see Fig. 2), our pipeline moves the view to the 2D representation of Fig. 4. We define the interaction region by two straight lines drawn at each time instant from the extreme points of the vehicle, thus the bottom-left point and top-right point. The two lines are parallel to the speed diference vector between the pedestrian and the vehicle ( ∆ ) and tangent to the contour of the vehicle. Upon detecting a potential conflict, the minimum distance between the two road users along the relative velocity ∆ is computed, followed by the calculation of the ITTC according to Equation 2. An example of such a conflict scenario is shown in Fig.4a, where the pedestrian lies within the lines defined by the vehicle’s contours.
This rigorous evaluation of the TTC measure instant-by-instant can produce gaps in the ITTC curve due to evasive actions, such as the pedestrian hesitations or the vehicle changes in direction or speed. Given a threshold of interest ℎ ( ℎ = 7 in our study) to filter the portion of interaction in which the actors are closer, we denote Sum of no-Interaction Times (SUM-no-IT) with Equation 3: SUM-no-IT = ∑︁ (no-IT) , when ITTC < Thrgap (3)
On the contrary, in Fig. 4b the road users are not in a potential collision course since the pedestrian is outside the interaction region defined by the two lines. This is due to the change in the relative velocities, which afects the slope of ∆ . This procedure provides a convenient mechanism to investigate the evolution of the whole interaction depending on the changes in speed and direction, thus identifying the most dangerous instants.
Pedestrian hesitation often results from evasive manoeuvrers or uncertainty in interactions with vehicles. This study examines pedestrian behaviour when crossing in mixed trafic, focusing on decision-making diferences between HVs and AVs. We identify two key events based on pedestrian speed profiles [28, 29]: • Stop Event: Occurs when the speed of the pedestrian drops below ℎ = 0.3 m / s, indicating hesitation (e.g. perceived risks or uncertainty about the intention of the vehicle). • Long Stop Event: A prolonged stop exceeding ℎ_ = 1 s, suggesting a deliberate delay in crossing.
These thresholds, derived from literature [30, 31] and exploratory analysis, efectively capture pedestrian decision-making dynamics and manage small fluctuations in the speed estimation. Two metrics quantify these behaviors: (i) Total Stop Time: sum of all interruptions during a crossing; (ii) Number of Long Stops: count of prolonged stops. These measures provide insights into pedestrian confidence and risk perception when interacting with AVs and HVs.
The study analyzed pedestrian-vehicle interactions at a non-signalized crosswalk along an autonomous shuttle (SAE level 4) route in an Italian city near a hospital. Video data were collected over 24 hours across eight days, focusing on one lane of a two-lane urban road to ensure data quality. A total of 168 interactions were examined, including 33 involving autonomous shuttles and 135 with human-operated vehicles, with all videos anonymized before processing.
The findings reveal that AV-pedestrian interactions were largely non-critical since they were predominantly within the no conflict range, with an ITTC mean value = 3.66 , = 2.278 . Even the lowest recorded ITTC for AVs remained above the serious conflict threshold of 1.5s, indicating a complete absence of high-risk interactions. In contrast, HV-pedestrian interactions frequently involved actual conflicts, as their average ITTC fell below the 3s ( = 2.885 , = 1.306 ). The most critical case of HV recorded a significant low ITTC of 0.331 s, indicating a serious safety risk. Additionally, AVs consistently prioritized pedestrian right-of-way, yielding 100% of the time, whereas HVs failed to yield in 20% of cases, highlighting an important contrast in vehicle behavior. (a) (b)
To compare pedestrian interactions with AVs and HVs, we analyzed the cumulative density functions (CDF) of ITTC and applied the Kolmogorov-Smirnov (K-S) test. The test did not show significant diferences between the two distributions ( = 0.168, = .415), which means that they share similar characteristics. We identified the lognormal distribution as the best fit for both AV-pedestrian ( = 0.076, = .985) and HV-pedestrian ( = 0.110, = .087) interactions. However, Fig. 5 revealed key distinctions: AVs consistently resulted in higher ITTC values, while HV interactions exhibited a more pronounced lower tail. Since lower ITTC values correspond to more dangerous interactions, the presence of critical conflicts in the HV distribution is concerning. In particular, AV interactions did not include such high-risk events, reinforcing their safer operational behavior.
Since TTC depends on both pedestrian and vehicle speeds, the observed time gaps between interactions result from their combined dynamics. However, pedestrian-HV interactions exhibit longer average durations ( = 1.55 , = 1.48 ) compared to pedestrian-AV interactions ( = 0.94 , = 0.75), with greater variability and higher maximum values. The highest recorded gaps ( = 6.53 for HV and = 2.31 for AV) can suggest evasive actions. To better understand pedestrian decision-making, we analyze stop patterns and speed variations during crossings (see Section 2.2.3). Stops, especially at the start of crossings, indicate hesitation and uncertainty. Pedestrians interacting with AVs tend to have shorter and more consistent stop durations, reflected in a narrower interquartile range ( = 0.204 ) compared to HV interactions, which show greater variability ( = 0.99 ) and extreme outliers up to 5.7 s. Also, Long Stops (1 s) occur in 12.12% of AV interactions versus 23.53% for HVs. Only for HVs, we have observed cases with multiple long stops for the same interaction. This suggests that HVs, with their less predictable behavior, induce more hesitation; while AVs facilitate more confident and predictable pedestrian crossings.
As an example, Fig.6 captures a pedestrian’s decision-making process during a pedestrian-HV interaction, depicting ITTC evolution (Fig.6a) alongside speed variations (Fig. 6b). At the start of the crossing, two prolonged stops emerge, marking moments of hesitation as the pedestrian evaluates whether to proceed. These pauses correspond with extended no-Interaction Times, as seen in the ITTC curve, where the pedestrian either slows significantly or comes to a full stop, thus showing signs of uncertainty. This pattern underscores that the most critical moments of indecision occur at the beginning of crossings, when pedestrians assess potential risks. Long stops are not only afected by the pedestrian’s own judgment but also by the vehicle’s responsiveness. This dynamic is especially relevant at uncontrolled crossings, where yielding behavior is uncertain and influences pedestrian confidence in proceeding.
The introduction of AVs in mixed trafic must be carefully evaluated to ensure safety. This study analyzed real-world pedestrian-vehicle interactions, demonstrating that AVs exhibit a lower collision risk than HVs with safer interactions at unsignalized crosswalks. Using the proposed pipeline to build a rigorous top-view 2D representation of the scene, we extracted spatial-temporal trajectories and computed conflict measures, revealing safer interactions with AVs, which consistently yielded to pedestrians with respect to HVs. Pedestrian behavior varied depending on vehicle type: interactions with HVs involved longer hesitation and stops, while AVs’ predictable behavior fostered smoother crossings. However, the conservative approach of the AVs, including waiting until the crosswalk was completely clear, can impact trafic flow. AV settings should be optimized to balance safety and eficiency, adapting speed and acceleration based on local trafic dynamics. Our findings provide valuable insights for AV integration, ofering data for micro-simulations and informing both transport operators and AV providers on optimal deployment strategies.
Future research could extend this study to more complex crosswalk designs, incorporating factors like multiple lanes, median refuges, or advanced warning systems. Additionally, the proposed methodology could be applied to a broader range of interactions involving vulnerable road users and diverse environmental conditions, such as nighttime crossings or even diferent countries, to enhance its generalizability. As urban mobility shifts towards eco-friendly modes like bikes and scooters, their distinct trafic dynamics introduce new safety challenges. Investigating interactions between autonomous shuttles and these road users could provide valuable insights into the risks and benefits of smart mobility solutions.
This work was partially supported by the funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement No 875530 (SHOW); and partially supported by the SmartData@PoliTO center on Big Data and Data Science.
The author(s) have not employed any Generative AI tools. [5] M. Wu, H. Jiang, C.-A. Tan, Automated parking space allocation during transition with both human-operated and autonomous vehicles, Applied Sciences 11 (2021). URL: https://www.mdpi. com/2076-3417/11/2/855. doi:10.3390/app11020855. [6] R. Gasper, S. Beutelschieß, M. Krumnow, L. Simon, Z. Baksa, J. Schwarzer, Simulation of autonomous roboshuttles in shared space, EPiC Series in Engineering 2 (2018) 183–193. [7] A. Tarko, Measuring road safety with surrogate events, Elsevier, 2019. [8] L. Zheng, T. Sayed, F. Mannering, Modeling trafic conflicts for use in road safety analysis: A review of analytic methods and future directions, Analytic Methods in Accident Research 29 (2021) 100142. URL: https://www.sciencedirect.com/science/article/pii/S2213665720300324. doi:https: //doi.org/10.1016/j.amar.2020.100142. [9] A. Laureshyn, T. De Ceunynck, C. Karlsson, Åse Svensson, S. Daniels, In search of the severity dimension of trafic events: Extended delta-v as a trafic conflict indicator, Accident Analysis & Prevention 98 (2017) 46–56. URL: https://www.sciencedirect.com/science/article/pii/ S0001457516303566. doi:https://doi.org/10.1016/j.aap.2016.09.026. [10] A. Papadoulis, M. Quddus, M. Imprialou, Evaluating the safety impact of connected and autonomous vehicles on motorways, Accident Analysis & Prevention 124 (2019) 12–22. URL: https://www.sciencedirect.com/science/article/pii/S0001457518306018. doi:https://doi.org/ 10.1016/j.aap.2018.12.019. [11] A. H. Kalantari, Y. Yang, J. Garcia de Pedro, Y. M. Lee, A. Horrobin, A. Solernou, C. Holmes, N. Merat, G. Markkula, Who goes first? a distributed simulator study of vehicle–pedestrian interaction, Accident Analysis & Prevention 186 (2023) 107050. URL: https://www.sciencedirect.com/science/ article/pii/S0001457523000970. doi:https://doi.org/10.1016/j.aap.2023.107050. [12] E. Beauchamp, N. Saunier, M.-S. Cloutier, "Study of automated shuttle interactions in city trafic using surrogate measures of safety", Elsevier, Transportation Research Part C 135 (2022). URL: https://doi.org/10.1016/j.trc.2021.103465. [13] D. Yang, K. Ozbay, K. Xie, H. Yang, F. Zuo, D. Sha, Proactive safety monitoring: A functional approach to detect safety-related anomalies using unmanned aerial vehicle video data, Transportation Research Part C: Emerging Technologies 127 (2021) 103130. URL: https://www. sciencedirect.com/science/article/pii/S0968090X21001492. doi:https://doi.org/10.1016/j. trc.2021.103130. [14] R. Madigan, S. Nordhof, C. Fox, R. E. Amini, T. Louw, M. Wilbrink, N. Merat, "Understanding interaction between Automated Road Transport System and other road users: A video analysis", Elsevier, Transportation Research Part F 66 (2019) 196–213. URL: http://doi.org/10.1016/j.trf.2019. 09.006. [15] A. Avignone, M. Bassani, B. Borgogno, B. Caroleo, S. Chiusano, F. Princiotto, Evaluating unsignalized crosswalk safety in the age of autonomous vehicles, Computers in Industry 167 (2025) 104259. URL: https://www.sciencedirect.com/science/article/pii/S0166361525000247. doi:https://doi.org/10.1016/j.compind.2025.104259. [16] J. Redmon, S. Divvala, R. Girshick, A. Farhadi, You only look once: Unified, real-time object detection, in: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 779–788. doi:10.1109/CVPR.2016.91. [17] T.-Y. Lin, M. Maire, S. Belongie, L. Bourdev, R. Girshick, J. Hays, P. Perona, D. Ramanan, C. L.
Zitnick, P. Dollár, Microsoft coco: Common objects in context, 2015. arXiv:1405.0312. [18] Y. Du, Z. Zhao, Y. Song, Y. Zhao, F. Su, T. Gong, H. Meng, Strongsort: Make deepsort great again, 2023. arXiv:2202.13514. [19] C. Hydén, The development of a method for trafic safety evaluation: The swedish trafic conflicts technique, Lund Institute of Technology, Department of Trafic Planning and Engineering (1987). [20] C. Hydén, Trafic conflicts technique: state-of-the-art, Trafic Safety Work with Video-processing,
Green Series, 43, University Kaiserlauten. Transportation Department (1996). [21] A. Tarko, "Measuring Road Safety with Surrogate Events", Elsevier (2019). [22] A. Arun, M. M. Haque, A. Bhaskar, S. Washington, T. Sayed, A systematic mapping review of surrogate safety assessment using trafic conflict techniques, Accident Analysis & Prevention