Study On Vehicle Environment Perception Method Based On Stereo Vision And Deep Learning

With the deepening of the research on vehicle intelligence and network interconnection,intelligent vehicle has gradually become the trend of development in the future.The key technology of intelligent vehicle lies in how to make the vehicle recognize the road environment accurately and timely.Therefore,it is of great significance to study a real-time and high accuracy vehicle environment perception algorithm in intelligent vehicle obstacle avoidance and automatic cruise.First,in order to solve the problem of long selection time and low recognition accuracy of obstacle candidate regions,an obstacle detection and recognition algorithm based on stereo vision and convolution neural network is proposed.The algorithm uses semi-global matching algorithm to calculate disparity map of left and right images,stixels calculation based on disparity map is to obtain candidate regions of obstacles,target obstacles are extracted from obstacles candidate regions based on U disparity map.In order to further recognize the extracted target obstacles,an improved convolution neural network based on AlexNet network is proposed.The experimental results show that the proposed convolutional neural network has better real-time performance and recognition accuracy than other networks.Then,aiming at the low accuracy of stixel representation in traffic scenes,a stixel segmentation algorithm based on deep learning is proposed.The algorithm adds semantic segmentation information on the basis of stixel representation,the information of ground,object and sky in traffic scene is extracted based on the disparity map acquired by stereo matching,and then the semantic segmentation information is generated by the semantic segmentation network ENet.The experimental results show that the accuracy of traffic scene representation is improved by adding additional semantics segmentation information on the basis of stixel representation.In addition,compared with other network models,ENet network has better real-time performance on the basis of maintaining better semantic segmentation accuracy,which lays a foundation for the application of the algorithm in vehicle equipment.Finally,in order to verify the performance of the algorithm in real traffic scenarios,a hardware experimental platform including TX2,ZED camera and other hardware devices is designed.The experimental scene is in Dalian University of Technology.The experimental results show that the algorithm designed in this paper has better effect of obstacle detection and recognition and traffic scene representation in actual traffic scene.