Deep Neural Network-Based Vehicle Longitudinal Control Using End-to-End Imitation Learning Fused with Logical Rules

Deep neural network- (DNN-) and imitation learning-based end-to-end visual solutions for autonomous driving have been of great interest to academia and industry. Compared to the rapid development of the essentially more vision-triggered vehicle lateral control based on end-to-end imitation learning, imitating a longitudinal control behavior according to the visual scenes in an end-to-end way, however, is still an active research topic. Towards this purpose, in this paper, an end-to-end imitation learning- and vision-based DNN model is developed for an Adaptive cruise control- (ACC-) like vehicle longitudinal control on the highway. Benefiting from its novel architecture to fuse learning- and rule-based mechanisms, in the simulation tests, the proposed DNN model achieves a satisfactory dynamic performance of the ACC-like longitudinal control with a high cruise control performance given a certain speed limit and a high vision-triggered speed adaptation performance facing a preceding vehicle, and outperforms the existing end-to-end imitation learning-based longitudinal control DNN model designed without fusing rule-based mechanisms.