Research On Lane Line And Target Detection Based On Deep Learning

Lane line detection and road target detection are important components of intelligent assisted driverless driving,and have important scientific research and engineering practice significance in the fields of intelligent transportation and unmanned driving.Due to the complexity of the road lanes and the variety of targets,the detection is also susceptible to environmental impact,which is a very challenging problem for lane lines and target detection.Therefore,this paper focuses on the research of lane line detection and target detection,which has certain theoretical and practical significance for intelligent assisted driving.After summarizing the current domestic and international research status of lane line and target detection,this paper clarifies the research content and objectives of this paper,analyzes and selects the convolutional neural network of this paper,which is convenient for subsequent lane line detection and target detection.Firstly,according to the traditional lane detection algorithm,based on the difference of the driver's field of view,a method of delineating the region of interest by the three-point field method is proposed and the lane pixels are detected and filtered and then the progressive scan threshold is used.The pixels are fitted.The analysis results show that the proposed algorithm has higher detection accuracy.Secondly,based on the deep learning theory,a dual-branch network and a custom function network are proposed to transform the lane detection problem into an instance segmentation problem.Firstly analyze the classical convolutional neural network,improve the encoder structure of the convolutional neural network convolutional layer,reduce the shared encoder parameters,and form a bifurcated network;then,end-to-end training to generate a convolutional neural network with a custom loss function.The parameter transformable matrix solves the problem of the pixel fit and difference of the far road and the road surface of the slope.By comparing the traditional detection algorithm with the results of classical convolutional neural networks,it can be found that the proposed combined network has higher detection accuracy.Finally,for the current monocular camera can not solve the target ranging and the target detection problem when the vehicle is occluded,we collect the calibration of the data set and finish the camera matching work.Mean while,the Mask R-CNN network and the YOLO network are used,and the binocular vision is used to complete the picture,video and inspection of real roads.The trial results of the enterprise show that the network can detect and track the target efficiently and quickly according to the input information,which has high adaptability and effectiveness.