This PhD thesis treats the “prediction of vehicle-to-pedestrian accidents” as part of a pedestrian precrash protection system. The study began with a thorough evaluation of these accidents. The data obtained from this work were then used to generate reference cases that allow the evaluation of intelligent pedestrian protection systems based on simulations. In the next step, the behavior of pedestrians was studied and a continuous-time statistical model, based on four discrete states, was proposed and evaluated. This pedestrian model makes possible a probabilistic prediction of vehicle-to-pedestrian accidents, for which Monte Carlo simulations were used. Variance reduction methods such as “splitting” and “Russian roulette” were used to improve the algorithms' performance. Finally, several hypotheses were proposed with aim of modeling the uncertainties involved in detecting pedestrians and estimating their relative positions and speeds system. Evaluating different types of uncertainties on these measurements, which are the inputs of the prediction module, contributes to the specification of a system that detects pedestrians and estimates their positions and speeds. This work led to three published articles and five patent applications.
