Pedestrian Avoidance with and Without Incoming Traffic by Using Deep Reinforcement Learning

Pedestrian avoidance is one of the most challenging autonomous driving operations in the field of intelligent vehicles. In an emergency, the optimal maneuver is to steer to avoid pedestrians and other vehicles. In this paper, a deep reinforcement learning based method has been proposed, in which the agent is trained to maneuver the ego vehicle steering away from the pedestrian with a safety clearance to the adjacent lane. A scenario both with and without other traffic have been investigated. By using TensorFlow as the learning framework and Unity3D to model the environment and different scenarios, the agent has been trained to obtain the maximum reward and take optimal policy in the process of continuously interacting with the environment. The capsule tool in Unity can ensure that the agent would keep the ego vehicle a safe distance from pedestrians during the training. The success rate of the trained agent in different conditions have proven the effectiveness of the proposed approach.