Lane Line Detection And Tracking Method Based On Deeplab-LatNet Network And Density Clustering

Lane line detection and tracking contribute to autonomous driving and vehicle deviation warning,which has been a hot and difficult topic in the automotive industry.The vehicle deviation can prevent vehicle trajectory deviation and reduce the occurrence of traffic accidents by monitoring the vehicle route.Fast and effective lane line detection and tracking method is an important basis to ensure driving safety.In the actual road scenes,the accuracy and stability of lane line detection base on image processing are face with greatly challenged,due to the influence factors,such as weather,light,buildings on both sides of the road,and the degradation of line signs.Therefore,in view of the above problems,this paper makes an in-depth research from the following three aspects: lane line image semantic segmentation,lane line fitting,and lane line tracking.The main work of the paper is as follows:(1)Aiming at the spatial position characteristics of lane line image in unban traffic scene,a longitudinal attention structure(LatNet)module was proposed to extract the lane line features.The LatNet model takes the low-level feature map as input and performs pooling operation in the horizontal direction,and finally outputs the weight matrix through the Sigmoid function after passing through the convolutional layer and fully connection layer.The proposed method can increase the ability of the model to extract the lane line structure in the vertical direction.(2)The Deeplab V3 + network was used as the backbone,the weight matrix generated by the attention structure module was fused into the feature map to obtain the final feature maps with weight characteristics.The fusion feature maps were segmented by semantic information to obtain the lane line segmentation image result.The comparative experiments were tested on public data sets to verify the feasibility and stability of the proposed model.(3)A density clustering algorithm was studied to cluster the lane line feature points in the segmentation graph,which can describe of the lane line features more effectively and eliminate the misclassification of pixel points in the semantic segmentation network partly.Aiming at the problems of current lane confirmation and structure disappearance,the leastsquare method was used to carry out linear fitting merging of current lane,and the Kalman filter was applied to predict and track the current lane line.