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

This work presents an approach to assessing the risk of interaction between a vehicle and a pedestrian, based on a combination of spatial, temporal, and dynamic information extracted from video images. Pedestrians are detected using a visual detection model, while the distance between them and the vehicle is estimated based on the geometry of a monocular camera. The speed of the vehicle is deduced from optical tracking based on the Lucas-Kanade algorithm. Using data from the World Health Organization on the severity of impacts, we propose a probabilistic formulation that allows us to jointly estimate the probability of collision and the potential severity of injuries. Unlike traditional indicators, which are limited to instantaneous measurements (speed, distance, or time to collision), our model takes into account the cumulative duration during which a pedestrian remains below a critical proximity threshold. This method of integrating temporal exposure provides a more accurate picture of the gradual evolution of danger. Experiments conducted on the JAAD dataset show that this approach allows for a more refined distinction between different levels of risk, particularly in urban situations at intermediate speeds where preventive actions remain possible. The results as a whole highlight the value of simultaneously integrating distance, speed, and exposure time for a more realistic and operational assessment of pedestrian-vehicle risk.