Deep Reinforcement Learning-Based Decision Making for Autonomous Vehicles under Diverse Pedestrian Crossing Scenarios

In the context of the rapid development of automated driving technologies, pedestrians, as the most vulnerable part of the environment among other traffic participants, should be given priority consideration to ensure the safety of pedestrians crossing the road. In this paper, an interaction strategy based on deep reinforcement learning is proposed for the safety of self-driving cars interacting with pedestrians. The strategy is based on deep reinforcement learning through the Deep Q Network (DQN) framework, which designs discrete reward functions for the sparse reward problem and the collision penalty problem regarding navigation, respectively, for the self-driving car’s behaviour of pulling out of the lane as well as pulling into the opposite lane, and a linear continuous reward function that combines the vehicle’s speed as well as the heading angle. Through experiments on the simulation platform CARLA, the algorithm achieved a 98% safe pass rate and improved efficiency by 41%.