Lane Detection Method Based On Deep Learning

With the development of intelligent transportation and environment awareness,Lane detection is one of the key technologies for Advanced Driver Assistance System and Autonomous Driving.With the development of deep learning,more and more algorithms based on deep learning have been applied to lane detection,but there are still many deficiencies in lane detection.Firstly,due to the simple appearance of lane lines and the lack of obvious features,other objects with similar local appearance can easily interfere with the detection process of lane lines.Secondly,the uncertainty of the number of lane and road change limits the accuracy and robustness of lane detection.Lane detection results are often unsatisfactory when there are shadows,degraded lane markings,and vehicles blocking the lane.In order to solve these problems,two methods of lane detection based on deep learning are proposed and compared.The first method is to design the ST-LanNet network,which consists of two parts.This network consists of two parts.The first part is the lane edge feature extraction network based on lane edge.In this part,the depth-separable convolution is used to replace the standard convolution in order to reduce the computation cost,and the lightweight general-purpose Content-Aware Re Assembly of Features(CARAFE)is used to restore the feature resolution to make the network lighter,finally,the proposed feature map of lane line edge is generated at pixel level.On the other hand,the traditional convolution-based lane location network is replaced by Swin Transformer in lane localization network to realize the precise lane location,and the proposed feature map of lane localization is obtained.At last,the lane detection network designed by us is tested by experiment.The results show that the lane detection network has good performance in dealing with difficult traffic scenarios and improves the accuracy and efficiency of the network detection.A large number of experiments based on two large-scale public datasets show that this method has obvious advantages over other methods in lane detection,especially when the vehicle is in a complex scene,lane detection is higher robust.The experiment of the method is based on TuSimple data set and Cu Lane dataset.The experiment results based on Tusimple dataset show that for easy scenarios,the validation accuracy is 97.46%,the test accuracy is 97.37%,and the accuracy is 0.865.For difficult scenarios,the validation accuracy is 97.38%,the test accuracy was 97.29%,the accuracy was 0.859,and the running time was 4.4 ms.The experiment based on Cu Lane data set shows that the validation accuracy is97.03%,the test accuracy is 96.84% and the accuracy is 0.837 for easy scene.For difficult scenarios,the validation accuracy is 96.18%,the test accuracy is 95.92%,the accuracy is 0.829,and the run time is 6.5 ms.In the second method,ST-MAE network is proposed to input six consecutive image sequences into the network.Through ST-LSTM block,the continuous multi-frame driving scene is modeled through time series,and the encoder composed of Swin Transformer block extracts the image information of each frame driving scene.Finally,through the decoder composed of Swin Transformer blocks,features are obtained and reconstructed to complete the detection task.A large number of experiments on two large-scale data sets show that the proposed method is superior to the competitive method in lane detection,especially when dealing with difficult situations.The experiment is based on TuSimple dataset.The results show that for simple scenarios,the verification accuracy is 97.46%,the test accuracy is97.37%,and the accuracy is 0.865.For complex scenes,the verification accuracy is 97.38%,the test accuracy is 97.29%,and the accuracy is 0.859.The running time is 4.4ms.The experiment is based on the CULane dataset.The results show that for simple scenarios,the verification accuracy is 97.03%,the test accuracy is96.84%,and the accuracy is 0.837.For complex scenes,the verification accuracy is 96.18%,the test accuracy is 95.92%,and the accuracy is 0.829.The running time is 6.5ms.