Vehicle control dynamic imitation learning algorithm based on surround vision

Abstract

The invention provides a vehicle control dynamic simulation learning algorithm based on surround vision and aims to solve the problem that a current condition simulation learning model in the field of automatic driving is poor in performance in a dynamic environment. According to the model, firstly, image features of continuous four frames of forward images are extracted by using a residual network, and then fusion feature vectors are obtained from the image features through an LSTM network. and dynamic environment feature vectors are obtained by combining the fusion feature vectors and side image features extracted by the residual network. Aiming at different navigation conditions, different decision networks are used for predicting the vehicle speed and the steering wheel angle, and finally a proportional integral control method is used for achieving vehicle longitudinal control. The DSCIL model successfully solves the problem of wrong mapping of low speed and low acceleration, the situation of low acceleration caused by low speed is completely eliminated; and the DSCIL model extracts a dynamic obstacle state by using multi-frame forward visual information, so that the performance of the model in a dynamic driving environment is improved.

Description

technical field [0001] The invention relates to the technical field of automatic driving, in particular to a dynamic imitation learning algorithm for vehicle control based on surround vision. Background technique [0002] In the process of human driving, it mainly relies on visual information to determine the traffic conditions ahead, relies on the speedometer to determine the motion state of the car, and relies on the neural network model to imitate human driving behavior. At the same time, human driving records and sensor information need to be collected for supervised learning. . In 2005, Lecun et al. built an end-to-end model DAVE with a 6-layer convolutional neural network, and used supervised learning to train the neural network. Research has shown that the model has good robustness in the wild environment. In 2016, NVIDIA trained a convolutional neural network model to predict the steering wheel angle by collecting real vehicle driving data. The feasibility of the e...

AI Technical Summary

Problems solved by technology

However, for the above models, it is difficult to obtain the motion state of dynamic obstacles, which is unimaginable in real driving

Images