Using External Attention in Vision-based Autonomous Driving

Imitation learning (IL) method provides a concise framework for autonomous driving, and it learns the policy from human demonstration via mapping from sensor data to vehicle controls. However, how to achieve effective learning is an ongoing challenge in driving policy learning. A feasible solution is to introduce attention mechanism, which enables the deep model focus on the features related to the driving task. In this paper, we propose to train a driving policy model with latest external attention joining the policy network as blocks. The experimental results on CARLA driving benchmark demonstrate our external attention guided driving policy model has better performance and costs less training time than the model without attention mechanism. Moreover, our method outperforms the state-of-the-arts self-attention driving policy methods.