Collision Avoidance of Autonomous Vehicles with E-bike at Un-signalized Occluded Intersections Based on Reinforcement Learning

Un-signalized occluded intersections are residential road intersections with narrow lanes and surrounding buildings, which are prone to traffic accidents. This work uses deep reinforcement learning to design driving strategies for Autonomous Vehicles (AVs) for avoiding collision and reducing damage to electric bicycles (e-bikes) with dangerous behaviors at un-signalized occluded intersections. The conflict-avoidance behavior of e-bikes is modeled. It adopts a multi-objective reward function that considers the injury severity of e-bike riders and the driving safety and comfort of AVs. A deep deterministic policy gradient method is used to train the model to control the acceleration and steering of AVs. The performance of the proposed method is compared with that of an autonomous emergency braking system and a risk-aware high-level decision strategy by simulation experiments. Experimental results show that the driving strategy can reduce the collision probability by 26.38% on average, and the injury can be reduced by 14.05% on average when the collision is unavoidable. To our knowledge, this is the first paper that employs reinforcement learning to model and design driving strategies for AVs conflicting with e-bikes. It can be used to improve the state of the art in AV control and safety at intersections.