Cooperative Decision-Making of Connected and Autonomous Vehicles in an Emergency

Abstract

Safety is one of the major concerns in autonomous driving tasks, and enhancing the collision avoidance ability of connected and autonomous vehicles (CAVs) is an effective way to improve road safety. Most current autonomous driving algorithms make braking and stopping decisions in a traffic emergency. However, such decisions may not be optimal. In this paper, we study the optimal collision avoidance decision-making method for emergencies. To address this challenge, we propose a cooperative decision-making scheme for CAVs in an emergency. Unlike previous decision-making methods, our scheme enables vehicles to avoid collisions by indicating the optimal emergency destinations, which leads to a new task: the prediction of the optimal collision avoidance destination. In the proposed scheme, the potential collision avoidance destinations are evaluated based on a deep reinforcement learning (DRL) model, and we propose a safety evaluation map (SEM) to describe the evaluation results. The cooperative ability of CAVs is considered in the proposed scheme, and a well-designed reward function is applied to train the DRL model. Extensive experiments demonstrate that the proposed model can accurately evaluate potential collision avoidance destinations and is effective in reducing traffic accident rates and accident damage in various traffic emergencies compared to state-of-the-art baseline methods.