A Data-Driven Spatio-Temporal Video-Based Model for Pedestrian-Vehicle Risk Prediction

BENKHADDA, Oumaima; MANDAR, Meriem; MELLOULI, Nédra · 2026 · Crossref

DOI: 10.21203/rs.3.rs-8743030/v1

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the limitations of existing pedestrian-vehicle risk assessment models, which typically rely on instantaneous metrics like speed or distance and fail to capture the cumulative nature of hazard exposure. The authors propose a data-driven, spatio-temporal model that integrates vehicle speed, pedestrian-vehicle distance, and cumulative exposure time to provide a more realistic and operational assessment of collision risk and injury severity. The motivation stems from the need for real-time, video-based monitoring systems that can distinguish between varying levels of danger in dynamic urban environments, particularly where preventive actions are still possible. The methodology combines computer vision techniques with kinematic analysis calibrated using World Health Organization (WHO) data. Pedestrians are detected using the YOLOv11 model, while the distance between the vehicle and pedestrian is estimated via monocular camera geometry, assuming standard pedestrian heights. Vehicle speed is derived from the Lucas-Kanade optical flow algorithm, which tracks static scene features across video frames. The risk model is probabilistic, defining risk as the product of collision probability and injury severity. Severity is modeled as a non-linear function of speed, calibrated to WHO statistics indicating a 4.5-fold increase in fatality risk between 50 km/h and 65 km/h. Probability is modeled using a logistic function that incorporates speed, distance, and the cumulative time a pedestrian remains within a critical proximity threshold of five meters. Experiments were conducted on the JAAD dataset, which contains 346 high-resolution video clips of urban driving scenarios. The study compared the proposed enriched model against a simplified model relying solely on speed. Results indicate that the enriched model provides a more refined distinction between risk levels, particularly in the intermediate speed range of 40 to 80 km/h. In this range, the enriched model systematically yields higher probability estimates than the speed-only model, reflecting the amplified risk caused by sustained proximity to pedestrians. At speeds below 30 km/h, both models predict low risk, while at speeds exceeding 90 km/h, both converge to near-certainty of high risk, rendering the additional spatial and temporal variables less distinct. The distance estimation proved robust to variations in assumed pedestrian height, and the optical flow method accurately captured acceleration, deceleration, and stationary phases. The significance of this work lies in its demonstration that integrating cumulative exposure time and spatial proximity alongside speed offers a superior framework for pedestrian safety assessment. By moving beyond instantaneous measurements, the model captures the gradual evolution of danger, providing more accurate alerts for preventive measures. This approach supports the development of advanced driver assistance systems and autonomous vehicle technologies capable of real-time risk evaluation in complex urban traffic conditions.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• telematics crash prediction