Research On The Lane Detection And Classification Method In Complex Scenes

In actual driving scenarios,the different types of lanes have different indication functions.It is very crucial for smart cars to detect and recognize the type of lanes accurately and quickly.This thesis designs a real-time lane classification detection network,which is composed of a lane detection network and a lane classification network cascaded,which can accurately and quickly complete the lane detection and classification tasks in complex scenes.Ⅰ.Lane detection is defined as a position selection and a classification problem in the direction of the row and a fast lane detection method based on row anchor is designed,which greatly improves the speed of detection.Since the width of the lane changes at different scales and there is a strong spatial relationship between different positions of the lane,the Res2 Net module is applied in the feature extraction network to extract multi-scale feature maps and coordinate attention is introduced in the feature extraction network to solve the long-distance dependence of lanes.Aiming at the low detection accuracy of lanes in curve scenarios,an auxiliary branch of vanishing point detection is added to assist in identifying the position and direction of lanes,improving the accuracy of detection in complex road conditions such as curves.The proposed method is verified on the Tu Simple and the detection accuracy of the proposed method is up to 96.78%,and the speed reaches 216.86 FPS.Ⅱ.A classification network LCNet based on lightweight convolution is designed,which improves the speed of the classification network while ensuring classification accuracy.Based on the results of the lane detection network,two approaches are presented to generate descriptors as input to the classification network.To solve the problem of the low classification accuracy and the poor efficiency of the HNet,a lightweight dual-feature fusion module DFF based on depth-separable convolution and inverted residual structure is designed,and the dual-feature fusion method is used to fully integrate the shallow features and the deep features of the network.A feature aggregation module FCT based on pooling operation in the downsampling layer is proposed,and through fusing the results of maximum pooling and minimum pooling,more original features of the image are preserved and the classification accuracy of the network is improved.Experiments show that the speed of the lightweight classification network proposed in this thesis can be up to 237.66 FPS,and the accuracy rate can reach 92.98%.The proposed method is 2.3 times faster than the Cascaded-CNN method.