Dynamic Imitation Learning Algorithm Based on Surrounding Vision

Abstract

Aiming at the poor performance of current conditional imitation learning model in the field of autonomous driving in dynamic environment, a dynamic conditional imitation learning model using LSTM network to fuse historical visual information is proposed. The model first extracts the image features from the front images of four consecutive frames by using the residual network, and then obtains the fusion eigenvector through the LSTM network. The dynamic environment eigenvector is obtained by fusing eigenvector and the side image feature extracted by residual network. Then according to different navigation conditions, different decision networks are used to predict vehicle speed and steering wheel angle. Finally, proportional integral control method is used to realize vehicle longitudinal control. The experimental results show that the vehicle control can be better.