Research And Simulation Of End-to-end Autopilot Model Based On Deep Learning

Auto-driving can effectively alleviate traffic safety accidents,improve traffic efficiency and greatly improve the existing traffic environment.At the same time,the landing of auto-driving is also an important part of the new national infrastructure,so the development of related technology has important practical significance.At present,there are two kinds of auto-driving schemes.Rule-based schemes need to design environment-aware algorithms manually,and complete a series of algorithms such as vehicle positioning,road strength planning,decision control,etc.with high-definition maps.The whole system design is complex and costly,and the whole system is not entirely impressive because of the ambiguity in defining the scenarios before them.End-to-end auto-driving scheme is relatively simple,the whole system only depends on the perceptual input of the vehicle,and the control output of the vehicle is obtained directly through the calculation of the model.Although the system is simple and low-cost,there are many problems such as the system is not visible,limited by data scenes,and there is no theoretical basis for system security.To solve the problems of system invisibility and data limitation,end-to-end scenarios are explored in combination with Neural Loop Policy Network(NCP)and deep reinforcement learning.The main contents of this paper are as follows:(1)The supervised learning scheme designs CNN+FCNN model and CNN+NCP model respectively.The function of the two models is the same except that they are different in structure,that is,they can predict the vehicle steering and speed at the same time.Based on Yushi technology’s industrial automatic driving simulator,data collection and model online simulation are completed.The two models are compared and analyzed from the four aspects of model training,model parameter quantity,model prediction effect and model processing speed.At the same time,the interpretability analysis of model feature extraction layer and mapping layer is completed.The experimental results show that the model based on NCP network has better robustness and generalization ability than the model based on fully connected network with less network parameters,and the interpretability of NCP network is also verified.(2)In the deep reinforcement learning scheme,behavior value network,state value network and strategy network are designed,and general reward function is also designed.Combined with lgsvl,which is the open source automatic driving simulator of Samsung and Google,the simulation training and verification are carried out.Two maps are trained in the experiment.The first map is relatively simple,and the experimental effect is obvious,and the model can converge quickly.In the second complex map,the agent falls into the local optimum and is difficult to jump out,and finally fails to achieve the desired goal.Although the scheme failed to achieve the desired effect,it verified the effectiveness of the reward function.Finally,it summarizes and analyzes the shortcomings and improvement direction of deep reinforcement learning scheme.