Study On Lane Line Detection In Complex Road Scenes Based On Deep Learning

The growing incidence of traffic accidents is a significant issue for contemporary society,and distracted and fatigued driving are the main contributors to high traffic deaths.For this reason,research on advanced driver assistance systems is developing rapidly.As its key components,lane departure warning systems and lane keeping assistance systems depend on the accurate detection of lane lines on the driving road,which aim to assist vehicles in quickly determining suitable driving areas and making the right decisions for drivers when the vehicle unintentionally drifts out of its lane,thus minimizing traffic accidents.The traditional lane line detection algorithms need manual adjustment of parameters,which not only has a large workload,but also has poor performance and generalization in the face of complex road scenes.In recent decades,the fully convolutional network based on deep learning has played an important role in the lane marking segmentation model,offering higher accuracy and speed compared to traditional methods.Although these networks are very efficient,it is still a very challenging task to use deep learning to reasonably determine the location of lane lines in complex driving scenarios like severe occlusion,extreme lighting condition,changing traffic flow,and poor road conditions.Additionally,the majority of deep learning models sacrifice lightweight network structures in exchange for excellent segmentation accuracy by adding a large number of modules to the network.Therefore,it is worthwhile to investigate how to use fewer parameters to obtain the accuracy and real-time performance of lane line detection algorithm under complex road conditions.To cope with the above problems,the following work is done in this paper:(1)To address the poor detection performance of existing lane line detection tasks in complex road scenes and the limitations of mainstream methods based on convolutional neural network(CNN)in acquiring global features of images,an axial-Edge Vi Ts-based dual attention enhancement feature fusion network(AEDANet)is proposed.Firstly,this model extracts the local and global features of lane lines in road images using a parallel coding branch structure that combines the strengths of CNN and Transformer in feature extraction.Secondly,the Transformer branch uses a highly lightweight network,Edge Vi Ts,and axial attention is used instead of self-attention to further reduce the computational complexity of the model.Finally,to make up for the features extracted at each stage,the coarse-grained features and fine-grained features are fused with each other using selective feature fusion modules and semantic guided channel attention modules to improve the performance.Experiments on the Tu Simple dataset show that AEDANet is superior to most advanced methods,which can better reduce error detection and enhance the lane line structure,and its F1-measure is as high as92.52%.(2)To solve the inadequacies of existing single-frame input models in extreme road scenes,a spatio-temporal information processing network for image key feature enhancement(SIPNet)is proposed.Firstly,based on the feature of great similarity between adjacent frames of road images,the model uses the feedback mechanism of Conv GRU on time dynamics to control the time series information.Secondly,the framework of CNN and Transformer is replaced with a tokenized shift multilayer perceptron block,which reduces the computational complexity of the model while improving the segmentation effect.Also,to further solve the problem of increasing the number of parameters caused by the simultaneous input of multiple frames,the regular convolution of the standard decoder is replaced with depthwise separable convolution.Finally,reverse attention mechanisms are employed throughout the downsampling process to reduce the impact of background noise and highlight the lane line features.Experiments on the Tu Simple dataset show that SIPNet outperforms the existing representative lane line detection models,and the F1-measure can reach up to 94.16%.Meanwhile,a comparative experiment with AEDANet on a complicated dataset of rural road indicates that SIPNet is more robust and can effectively detect lane lines in challenging driving scenarios.