Learning A Steering Decision Policy for End-to-End Control of Autonomous Vehicle

Steering control plays an important role in vehicle driving decisions. Convolutional Neural Network(CNN) has been widely studied in the field of autonomous vehicle navigation due to its powerful nonlinear expression and spatial feature understanding ability in image scene. In this paper, we present an end-to-end steering controller based on CNN to predict the desired steering wheel angle from a continuous video image, which does not require manual design rules and simplifies a series of intermediate steps in traditional autonomous driving decision, such as target detection, object recognition and path tracking. The pre-trained network model is verified in the TORCS racing simulator. The experimental results show that the controller has good generalization ability and can make the vehicle follow the right side of the lane on the unknown test track. In order to intuitively understand what factors in the image have an impact on vehicle decision-making, the remarkable features that affect the steering control of the autonomous vehicle are visualized.