Research On Lane Detection And Vehicle Behavior Prediction Algorithm In Complex Multilane Scenarios

Intelligent connected vehicles(ICVs)and autonomous driving technology have attracted significant attention from academia and industry,as they hold great potential for effectively addressing traffic safety and efficiency issues.In ICV technology,lane detection and vehicle behavior prediction are key techniques for achieving vehicle autonomy and driver assistance,and they have gradually become research hotspots.Vehicle behavior prediction technology enhances traffic safety,while high-precision lane detection serves as a crucial foundation for achieving accurate vehicle behavior prediction.However,existing lane detection and vehicle behavior prediction techniques are prone to interference and impact from various factors in complex multi-lane environments,thus failing to provide effective safety assurance.Therefore,this paper conducts in-depth research on lane detection and vehicle behavior prediction in complex multi-lane environments.The specific research focuses are as follows:(1)A self-adaptive lane detection method based on the You Only Look Once(YOLO)v5algorithm is designed.Modifications are made to the structural parameters of the one-stage YOLO v5 algorithm,constructing a two-stage learning network that is more suitable for lane detection.To adapt to different driving environments,an adaptive learning framework is incorporated into the two-stage learning network,allowing for adaptive learning of lane features in various environments.Additionally,to improve training efficiency,a method is proposed to automatically generate lane line label images in simple scenarios.Simulation experiments demonstrate that the model achieves detection accuracies of 91.7%,72.5%,and 78.3% in regular,nighttime,and congested environments,respectively.(2)A lane detection and prediction fusion algorithm based on the Dempster-Shafer(D-S)evidence theory is established.To address the low detection accuracy of lane detection based on the YOLO v5 algorithm in challenging environments such as occlusion,deformation,and wear,a lane prediction model based on convolutional neural networks,bidirectional long shortterm memory networks,and attention mechanisms is proposed.This model takes the lane detection information outputted by the YOLO v5 algorithm as input to predict the missing lane positions.Furthermore,to further improve the accuracy of lane detection and prediction,the DS evidence theory is utilized to fuse the prediction model and the lane information outputted by the YOLO v5 algorithm.In simulation experiments,the proposed fusion algorithm enhances the lane detection accuracy by 2%-3%,indicating that the fusion algorithm effectively improves the accuracy of lane detection.(3)A vehicle behavior prediction model is constructed based on the hierarchical Support Vector Machine(SVM)and Gaussian Mixture Hidden Markov Model(GM-HMM).SVM is utilized as a first-stage vehicle behavior classification and prediction model,with predicted outcomes including lane keeping and lane changing.When the prediction is lane keeping,the result is directly outputted by SVM.When the prediction is a lane-changing behavior,the model enters the second stage,employing the GM-HMM model to further predict the specific lane change direction(left or right).Additionally,a lane change data extraction algorithm is employed to enhance the prediction accuracy of the model.In simulation experiments using the High D dataset created by the Technical University of Berlin,the model achieves an accuracy of approximately 93% in vehicle behavior prediction.This represents an improvement of around 3%,4%,and 4% compared to traditional Hidden Markov Models(HMM),SVM,and multi-layer perceptron models,respectively.