End-to-end Autonomous Driving Method Based On Spatiotemporal Neural Network Model

Autonomous driving technology is of great significance to the development of automobiles and transportation.In recent years,with the development of core technologies such as artificial intelligence and sensors,autonomous driving technology has become a research hotspot.At present,autonomous driving technology is still immature.The intelligent and autonomous decision-making control of automobiles is the research focus of autonomous driving technology.This thesis realizes the autonomous decision control of the vehicle by studying the end-to-end automatic driving method.The end-to-end automatic driving method makes decision control by learning the driver's driving experience through the end-to-end automatic driving algorithm model.Compared with the traditional automatic driving method,this method does not involve complex environmental awareness tasks,and does not require human-made rules to generate driving decisions,which can effectively reduce the complexity of the automatic driving system.This thesis improves the effect of the end-to-end automatic driving method by designing a high-performance end-to-end automatic driving algorithm model.The main works of this thesis are as follows:(1)This thesis designs a spatiotemporal end-to-end automatic driving algorithm model based on spatiotemporal neural network model.This thesis considers the continuity of the driving process and designs a spatiotemporal end-to-end automatic driving algorithm model that can make driving decisions based on current and historical road information.The model takes the road image sequence collected by the on-board camera as input,and can extract spatial and temporal features from the input and predict the vehicle's steering angle and acceleration / deceleration control.The experimental results show that,compared with the traditional end-to-end automatic driving algorithm model,the spatiotemporal end-to-end automatic driving algorithm model designed in this thesis predicts more accurate driving decisions and can better control the vehicle to complete the automatic driving task in the simulation environment.(2)This thesis designs an end-to-end automatic driving algorithm model of visual attention.Inspired by human visual attention,this thesis adds visual attention mechanism CBAM to the end-to-end automatic driving algorithm model to form an end-to-end automatic driving algorithm model of visual attention.CBAM can perform weighted calculation on the image feature maps during the calculation of the algorithm model,thereby focusing on the important information related to driving decisions in the road image.The experimental results show that,compared with the original algorithm model,the end-to-end automatic driving algorithm model with increased visual attention mechanism has lower error in predicting driving decisions on the test data set and performs better in the autonomous driving tasks based on the simulation environment.(3)This thesis visualizes and analyzes the end-to-end automatic driving algorithm model.In this thesis,feature graph visualization and model learning feature visualization are performed on the designed end-to-end automatic driving algorithm model.The visualization results show that the end-to-end automatic driving algorithm model can abstract the features in the road image.The features learned by the algorithm model are the lane lines and vehicles in the road image,that is,the model can recognize the lane lines and vehicles and make driving decisions based on environmental information.This result conforms to the correct driving logic,and proves the rationality of the end-to-end automatic driving algorithm model.There are 37 figures,5 tables and 60 references in this thesis.