Lane Detection Algorithm Based On Semantic Segmentation And Pixel Vector Clustering

Lane detection is a technology that can accurately and efficiently identify the lane line of the driving scene.It is the basic task of automatic driving.Research on lane detection is of great significance to lane keeping and vehicle following.In recent years,many lane detection models have been proposed.Efficient and stable lane detection in complex driving scenarios is still an open problem.This thesis proposes a lane detection algorithm based on high-Precision semantic segmentation network and an efficient and robust lane clustering algorithm for lane binary segmentation and lane instance segmentation,respectively,and integrates the two algorithms to achieve efficient and credible lane instance segmentation.The main research contents of this thesis are as follows :(1)A lane detection algorithm based on U-Net semantic segmentation network is proposed : U-Net network structure has strict symmetry,and can use different resolution feature maps for feature fusion to obtain high-Precision lane edge information;aiming at the complex driving scenes such as vehicle occlusion,lane line wear and bad weather,the spatial feature extraction module of Atrous Spatial Pyramid Pooling(ASPP)and the spatio-temporal feature extraction module of long short-term memory convolution neural network(Conv LSTM)are embedded on the basis of U-Net network,respectively.The lane line detection models of U-Net＿ASPP,U-Net＿Conv LSTM and U-Net＿ASPP＿Conv LSTM are constructed,and the performance verification of the models is completed on Tu Simple and rural road data sets.The index used to measure the stability of the binary segmentation model is F1＿score;performance verification results show that the F1＿score of U-Net＿ASPP model increases from 0.8128 of U-Net to 0.8374,the F1＿score of U-Net＿Conv LSTM model increases from 0.8128 of U-Net to 0.8409,and the F1＿score of U-Net＿ASPP＿Conv LSTM model increases from 0.8128 of U-Net to 0.8560.(2)A lane clustering algorithm based on pixel vector domain is proposed.The Affinity Field(AF)algorithm can assign horizontal and vertical vectors to each lane line pixel from the bottom to the top of the pixel row of the road video frame,and complete intra-class aggregation and inter-class segmentation through the distance threshold between vectors to realize lane line clustering.In addition,based on the pixel vector domain clustering algorithm,this thesis proposes a K-AF clustering optimization algorithm.This algorithm uses the slope change of the lane line pixel center point connection between each adjacent pixel rows to screen out the key rows that are greater than the custom threshold,and only assigns horizontal and vertical vectors to the pixel center point(highlight point)in the key rows,so as to realize the optimization of operation efficiency.(3)Lane detection algorithm based on high-Precision semantic segmentation and pixel vector clustering.This algorithm is the overall algorithm of lane line instance segmentation proposed in this thesis,which integrates lane line binary segmentation algorithm and pixel vector clustering algorithm.The binary segmentation algorithm can run independently,and the implementation of pixel vector clustering algorithm depends on the segmentation results of the binary segmentation algorithm.Then the U-Net＿ASPP＿Conv LSTM model is used to extract the spatio-temporal characteristics of the lane line with high Precision to obtain the binary segmentation results.The binary segmentation results provide pixel position information for the pixel vector clustering algorithm,and then the proposed K-AF algorithm is used to complete the lane line instance segmentation.The lane line detection synthesis algorithm proposed in this thesis can obtain the space-time information of lane line with higher accuracy,and performs well in the task of lane line binary segmentation;In close range pixel clustering,the branch of pixel vector clustering performs well;Without specifying the number of lane lines,the number of lane lines can be accurately predicted and the task of lane line instance segmentation can be completed.