Ventura R., Roussou S., Ziakopoulos A., Barabinoa B., Yannis G., “AI-Powered Tracking and Analysis of Pedestrian and Vehicle Behaviours with Time-to-Collision Estimation”, Proceedings of the 2025 Transportation Research Symposium (TRS), Rotterdam, 25-28 May 2025.

The integration of Computer Vision techniques has greatly advanced road safety analysis, particularly in assessing traffic conflicts and near-miss events, typically measured through Surrogate Safety Measures (SSM). Nonetheless, there are still gaps in the literature regarding integrating multiple advanced models, real-world coordinate mapping, and comprehensive behavioural analysis within a unified framework. This study addresses these challenges by introducing an advanced detection and tracking algorithm to monitor pedestrian and vehicle behaviours in complex urban environments, with a particular focus on interactions with traffic signals Time-to-Collision (TTC) estimation, and Post Encroachment Time (PET). The system achieves high precision and robustness by combining YOLO-v10, ResNet-50, and Re-ID object detection and identification models with Kalman filtering, homography transformations, and re-identification techniques.
Using street-level video footage from a specific day at a location in Athens, Greece, the discrepancy between automated and manual counts is compared with those derived from a previous detection algorithm integrating YOLOv8. The evaluation focuses on the relative performance of both methods in capturing illegal pedestrian and vehicle behaviours. Additionally, analysis of TTC and PET trends reveals a clear inverse relationship between these variables and vehicle speed indicating the possibility of collisions. Furthermore, illegal crossings are associated with consistently lower TTC and PET values, underscoring a heightened chance of near-collision incidents.
These findings underscore the effectiveness of the proposed approach, as demonstrated through a comparative analysis of error and accuracy metrics with alternative detection methods. This analysis highlights the superiority of our approach in accurately capturing high-risk behaviours, particularly illegal crossings, which are critical yet often neglected in traffic safety research. By focusing on illegal crossing behaviours, this study contributes a novel perspective to urban safety analysis, addressing a significant gap in the existing literature and providing valuable insights for data-driven interventions in traffic management and policy development.