Semantic Segmentation And Fitting Tracking Of Lane Lines In Complex Road Scenes

Lane detection and tracking technology is the core aspect of vehicle intelligent driving,providing accurate road information for safe vehicle driving.While there have been some advancements in semantic segmentation of lane line in recent years,limitations in complex road scenarios result in both missed detections and false detections.Additionally,research on lane classification detection has been insufficient,as traditional methods typically treat all lane markings as the same category,lacking traffic semantic information.The segmentation of different types of lane markings is more practical for intelligent driving.Therefore,this study proposes an optimized Seg Former semantic segmentation algorithm to improve the robustness and accuracy of different types of lane marking segmentation in complex road scenarios.The main objectives are as follows:(1)Data acquisition,pre-processing and labeling.The monocular camera is used to collect eight types of lane lines in complex road scenes such as lightweather,road wear,shadow,and shading,and the eight types of lane lines are single dashed white line,single dashed yellow line,single solid white line,single solid yellow line,double solid white line,double solid yellow line,double line-left dashed right solid,and double line-left solid right dashed,to establish a database of lane lines in complex road scenes and classify them accurately;To ensure the quality of the data,we perform hash de-duplication and Gaussian filtering to remove the noise;finally,the data is labeled with categories.(2)Research on lane line semantic segmentation algorithm.Firstly,Experimental analysis of Faster RCNN,FCN,OCRnet,U-Net and Seg Former semantic algorithms to select the optimal algorithm as the basis for optimization;Secondly,the Seg Former semantic segmentation algorithm is used as the base algorithm to embed a non-local neural network module to enrich the contextual information of lane line features and improve the segmentation accuracy of different categories of lane lines in complex scenes;Finally,RESA aggregation module is established for the lane line model of complex road scenes to enhance the global fusion of lane line category features in the encoding and decoding process,reducing the loss of lane marking features;experimental validation shows that: The MIo U of this model is 80%,which is 5% better than the base model,and the m PA is 91%,which is 5.4% better than the base algorithm.(3)Lane line fitting tracking based on key pixel points.For the problem of high time complexity in the semantic segmentation pixel point fitting process,a lane fitting method based on optimal key pixel points is proposed;Least square method are fitted to key pixel points;finally,the prediction update of lane lines is completed using Kalman filtering algorithm to ensure the stability of lane line tracking state while improving the timeliness of lane line fitting tracking.(4)Experimental analysis of complex road scenarios.The experimental results show that the lane line semantic segmentation network proposed in this study has strong robustness and accuracy in complex road scenarios compared with the basic model.The proposed lane line semantic segmentation algorithm can be applied to complex road scenarios,and the algorithm also provides more detailed and accurate classification of lane types,which provides more useful information for intelligent driving systems.The algorithm has greater robustness and higher accuracy than other lane segmentation algorithms.