Research On The Multi-Lane Detection Method In Complex Road Scene

By accelerating research on autonomous driving technology,people have improved the serious energy consumption and traffic congestion problems caused by the rapid popularization of cars.The lane detection is an important part of the automatic driving system.The detected lane is an important mark that restricts the driving of vehicles on the road and directly determines the driving control and path planning of the vehicle.Therefore,being able to accurately detect the lane in the road is of great significance for improving the driving safety of vehicles.The development of lane detection technology enables the application of automatic driving systems in simple road scenes.However,due to poor lighting conditions,occlusion by other traffic participants,interference from irrelevant road markings and the inherent shape of lanes,complex road scenes the detection of lanes presents challenges and hinders the commercial application of autonomous driving technology in urban road scenarios.This paper addresses the problem of lane detection in complex road scenarios above,and proposes a lane detection network based on object feature distillation and a lane detection network based on geometric attention-aware.The specific contents are as follows:(1)A general lane detection framework based on object feature distillation is proposed.First,in a network using direct upsampling,add a decoder with strong feature prediction capabilities.This decoder plays a key role in the restoration of lane features in complex road scenes;then,in the network training stage,pass Knowledge distillation technology uses the prediction results generated by the decoder as a soft target,so that the direct upsampling branch learns more detailed lane information,so that it has a stronger feature prediction ability of the decoder;finally,only need to use in the network inference stage the branch is directly upsampled without forward calculation of the decoder.Therefore,compared with the existing model,it does not increase the additional calculation cost and can also improve the lane detection performance.(2)A lane detection network based on geometric attention-aware is proposed.The network adopts a multi-task branch network structure.In addition to the lane segmentation branch,the distance embedding branch is added to learn the continuous representation of the distance from the center line of the lane to the boundary of the lane.The two branches are propagated through the adaptive selection of complementary information by the attention information propagation module,and at the end of the two task branches,a geometric attention-aware module is used for feature fusion,which converts the output feature of the embedded branch into an attention.Force matrix,which captures the long-distance correlation between pixels from the spatial dimension.These contexts containing distance information can effectively improve the results of semantic segmentation.Finally,using the skip pyramid fusion upsampling module,the abstract features of different lane in each level encoder of the geometric attention-aware network are fused with the network output features that restore the resolution during the upsampling process,which effectively improves the ability to detect lane boundaries.In order to verify the effectiveness of the lane detection network based on object feature distillation,the current challenging CULane dataset is evaluated.The experimental results show that the F1-Measure of the lane detection results in complex road scenes in the CULane dataset The scores have improved,and the total F1-Measure has reached 74.1%.In order to verify the versatility of the object feature distillation network,it is applied to multiple mainstream lane detection networks.The experimental results show that the lane detection capabilities have been improved on the CULane dataset.In order to verify the effectiveness of the lane detection network based on geometric attention perception,experiments were conducted on the CULane dataset,TuSimple dataset and BDD100 K dataset.The total F1-Measure of the geometric attention perception network on the CULane dataset The score reached 75.8%,Accuracy on the TuSimple dataset reached 96.75%,and the IoU score on the BDD100 K dataset reached 16.75.