Perception And Decision Of Self-Driving Based On Deep Reinforcement Learning

As the development direction of future traffic,driverless vehicles can effectively enhance road safety,alleviate traffic congestion,reduce air pollution and save more time for drivers.However,with the continuous improvement of the demand for automated driving,the driverless vehicle needs to identify and analyze complex road scenes,and make decisions for all kinds of unexpected situations.This poses a new challenge to the environmental perception ability,fast learning ability and generalization ability of the self driving vehicle.In this paper,a vision based solution is used to study the environment perception and decision-making of driverless vehicles in complex scenesFirstly,the lane detection algorithm in driverless environment perception is deeply studied,and a semantic segmentation multi lane detection method based on adaptive distillation learning is proposed.The neural network structure of light-weight vehicle lane detection is designed.The network structure is designed through the encoder decoder framework,and the multi feature fusion method(Concat)is added,so that the detail information of shallow features is introduced into the deep layer,so as to enhance the detailed features of lane edge and improve the efficiency of real vehicle lane detection.The multi-level adaptive distillation learning method is proposed to further optimize the algorithm In this paper,a YOLA branch is proposed to supplement the output of the neural network for the error detection of brake lines and shadows in specific scenes.Secondly,aiming at the target detection of driverless traffic scene,a target detection method based on one-stage network is proposed.By adding multi branch feature extraction structure,features are extracted and output in different receptive fields,which improves the performance of small object detection.A faster and more accurate frame regression method(LIo U)is proposed,which maximizes the proportion of detection frame intersection in the global,so as to increase the gradient,make the convergence faster and regression more accurate,and improve the level of target detection and location.Experiments show that the proposed traffic scene target detection algorithm has good detection performance in small object detection,and achieves good target positioning effect.Finally,aiming at the real complex road driverless decision-making problem,a ddpg driverless decision-making method based on supervised training feature network is proposed,which improves the efficiency of feature sample collection of reinforcement learning by obtaining feature extraction network in advance from imitation learning;a multi-source information fusion decision-making method is proposed,which integrates perceptual information into the decision-making system and designs a multi-source information fusion system The experimental results show that the algorithm can speed up the training speed and improve the stability of decisionmaking.