Research On End-to-End Autonomous Driving Decision Algorithm Based On Deep Learning

Autonomous driving technology is a research hotspot in the automotive field,which has a profound impact on the automotive industry and even the transportation industry.In the autonomous driving technology,the driving decision module is the core part,which is responsible for receiving the information from the driving perception module,and learning the scene information in the driving process through the algorithm,and output the appropriate steering angle,speed and other decision information to the control module,so as to complete the horizontal and vertical control of autonomous vehicles.This paper mainly studies the end-to-end autonomous driving decision algorithm.The end-to-end automatic driving method is to learn the driving experience of human drivers by establishing an efficient end-to-end algorithm model to output vehicle decision information.The end-to-end automatic driving algorithm model takes the images captured by the vehicle camera as the input,and outputs the decision-making information such as vehicle rotation angle and speed.Therefore,it avoids the complex environmental perception task of the perception module,greatly reduces the calculation amount and improves the speed of decision-making.This paper constructed an end-to-end multi-task decision-making model integrating spatio-temporal feature information and vehicle dynamics information,which can simultaneously realize the horizontal and vertical control of autonomous vehicles.The main work of this paper is as follows :(1)Aiming at the problem that it is difficult to obtain the end-to-end automatic driving data set that meets the requirements,this paper calls the vehicle model in the Carla driving simulator,sets the vehicle as the automatic driving mode to simulate human driving,and obtains the road image and vehicle dynamics information in front of the vehicle through the sensor to construct the simulation automatic driving data set.And real vehicle experiments will be carried out to obtain real data sets.(2)This paper studies the influence of different classical networks on the model decision performance when they are used as feature extraction networks of end-to-end driving decision algorithm,and selects the optimal classical network as the feature extraction network of end-to-end driving decision algorithm,and compares it with the traditional end-to-end driving decision algorithm Pilot Net.The comparison results show that the end-to-end driving decision algorithm studied in this paper has better performance in the prediction task of autonomous vehicle decision value.(3)In view of the problem that the general end-to-end driving decision algorithm does not use the spatio-temporal information between historical images and the single input information of the algorithm model,we first add the long and short-term memory network LSTM to the algorithm to combine with the feature extraction network to obtain the spatio-temporal information between image sequences,and then we will integrate the vehicle dynamics information in the algorithm to increase the input information of the model,so that the algorithm model can more smoothly and accurately predict the decision value of the vehicle.(4)Aiming at the problem that the key feature extraction ability of the end-to-end driving decision algorithm is not strong and the convergence speed of the model training is slow,this paper studies two improved methods: 1)Based on the improved method of transfer learning,the parameters of the feature extraction layer of the pretraining model are transferred to the end-to-end algorithm model,which greatly improves the convergence speed and stability of the model training;2)Based on the improved method of visual attention mechanism,the visual attention mechanism module is added to the end-to-end algorithm model,which enables the model to learn those features that play a key role in driving decisions and improve the feature extraction ability of the model.(5)This paper makes visual analysis and simulation experiments on end-to-end algorithm model,including model structure visualization and feature map visualization,which deepens the understanding of the internal structure of the model and reflects the significance of model feature extraction.The simulation experiment is carried out in the automatic driving simulator Carla to verify the effectiveness and feasibility of the end-to-end driving decision algorithm studied in this paper.The experimental results show that the end-to-end driving decision algorithm studied in this paper can well complete the automatic driving task of the test section.