| Lane detection plays an important role in the perception task of autonomous driving.However,lane detection is different from object detection in the general sense.Its structure is thin and long,showing a pattern that spans the entire image.In addition,the foreground visual features of lanes are very sparse compared to background pixels,and often encounter difficult driving scenarios such as lane degradation,occlusion,bright light,and rainy weather,resulting in the already sparse foreground visual features becoming even sparser and ultimately leading to poor lane line detection models.Based on lane information of different scales,we propose several solutions for these problems encountered in lane detection:1.First of all,for difficult driving scenarios such as the lanes of the current frame,is easily occluded and degraded in the automatic driving scene,this thesis proposes a convolutional GRU(ConvGRU)model to fuse the lane information of consecutive frames to assist the detection of the current frame.In addition,for the current lack of lane line datasets in complex scenes,this thesis generates four more challenging lane detection scene datasets through the style transfer algorithm to verify the robustness of different models.2.Secondly,the current lane detection methods based on semantic segmentation usually aggregate the context information of the lane itself in the vertical and horizontal directions of the feature map.The existence of this high-level semantic information is underutilized.This thesis designs a lane detection network based on spatial multiscale information aggregation,which consists of a multi-scale feature information aggregator(MFIA)and channel attention(CA).In MFIA,the current feature map can be enhanced by information aggregation networks of different scales,so that the feature map can communicate and share information in different spatial ranges;in CA,the high-level semantic information of the existence of lane lines can be It is fully utilized so that the network can capture the global semantic information of lanes.3.Finally,this thesis designs an HWTransformer model based on cross-row and crosscolumn,further considering the prior geometric characteristics of lanes.The selfattention in the HWTransformer model will be limited to each row and each column of the feature map,and each row and each column will have an intersecting feature,which can realize the sharing of different local information,so as to achieve the same level as the global self-attention model.same effect.In addition,we also proposes a plug-and-play self-distillation method for the Transformer model,which can stably improve the performance of the Transformer model without adding any training cost.We have conducted extensive experimental validation on widely used mainstream lane detection datasets,such as TRLane,TuSimple,CULane,VIL100,BDD100K,etc.,to demonstrate the robustness of the method proposed in this thesis in various difficult driving environments. |
