End-to-End Control Method for Unmanned Vehicle Based on Knowledge Distillation

Abstract

In order to construct a lightweight end-to-end autonomous driving model, knowledge distillation technology is introduced into the construction of the model. The model is composed of a vehicle trajectory orientation marker point prediction algorithm and a motion control algorithm. The prediction algorithm uses the Resnet34 with high precision as a teacher network and the Squeezenet with fast speed as a student network. Based on the knowledge distillation technology, the prediction ability of the teacher network is transferred to the student network so that the student network can predict a marker point accurately and quickly. The control algorithm is constructed based on the integral-differential model, and the predicted marker point is converted into lateral control commands of a vehicle. Through the experimental analysis of the complexity of the model, prediction accuracy of the marker point, and actual vehicle deployment effect, it is verified that the proposed model can control a vehicle to drive accurately and uniformly along a lane line trajectory.