Influence of different pedestrian behavior models on the performance assessment of autonomous emergency braking (AEB) systems via virtual simulation

Abstract

Pedestrian safety is a central topic in the automotive industry because of the high number of deaths in car-to-pedestrian accidents. Different systems have been developed to protect pedestrians and other vulnerable road users. So-called Active Safety Systems are used to avoid possible collisions with the VRU or to mitigate injury severity by reducing the collision speed in case the collision can no longer be prevented. The autonomous emergency braking system (AEB) is one of these systems and aims to intervene in conflict situations by stopping the car. The performance assessment of the AEB system can be done via virtual simulation. One crucial aspect is the modeling of pedestrian behavior. Current studies use a simple pedestrian behavior model, sometimes called a trajectory-based model, in which the pedestrian moves with constant speed on a given path and without any interaction with the environment. This study investigates how the AEB performance in virtual environments is influenced by using a more realistic pedestrian behavior model based on the reinforcement learning approach, a particular machine learning branch perfectly suited for modeling decision-making processes. For that, a generic AEB system, the trajectory-based pedestrian model, and the reinforcement learning model were implemented in the CARLA Simulator. A scenario catalog was created by varying some parameters and used to evaluate the front collisions with and without the AEB system. The study indicates that due to some pedestrian reactions of the reinforcement learning model, such as unexpected stopping in front of the car, the performance of the AEB system is reduced.