Incremental Automatic Vehicle Control Algorithm Based on Fast Pursuit Point Estimation

Abstract

Image-based autonomous driving control is one of the important research directions in the field of autonomous driving. Most of the existing image-based control algorithms use end-to-end mapping from image to vehicle control amount, which is not explanatory enough, and the control amount is not intuitive enough to effectively implement human-machine collaborative control and incremental learning of models. This paper proposes an incremental learning algorithm for driving vehicle control based on fast pursuit point estimation. We establish a model to calculate the mapping of image to the pursuit point, and then get the actual control amount of the vehicle throttle value and front-wheel rotation angle value by the pursuit point. Combining the features of pursuit point which can be observed intuitively and has obvious physical meaning, we propose an incremental model updating method based on man-machine collaborative control, which can incrementally improve the model performance in the actual driving process of vehicles. Finally, the experiment of automatic control is carried out on the Carla simulation platform. The experimental results show that the algorithm can incrementally improve the performance of the automatic control model, with the average calculation speed over 50fps. The autonomous driving system realizes automatic cruise in the real campus environment.