Vehicle Trajectory Learning Method For Intersection Conflict Analysis

With the rapid development of urban construction and the improvement of urban infrastructure facilities,the traffic video surveillance systems are deployed to every corner of the city.The access to video data streams greatly enhances the city's management level,enabling city managers to respond quickly to sudden incidents on the road and save people's property.For the work related to the security analysis of intersections based on video data,the following shortcomings still exist in the current research: Firstly,a large number of related research still relies on manual work,and people assess the safety of intersections according to some certain rules,which is not only inefficient but also labor costly.Secondly,the related researches are based on the simulation data,and proposes a vehicle behavior conflict analysis algorithm based on the motion model.Some of them only focus on the feasibility of the behavior verification algorithm between several real vehicle trajectories,and ignore the complexity of the scene and the resulting low quality data problems.Thirdly,the behavior analysis based on vehicle trajectories has a lot of related work,and for the current research on safety,vehicle behavior analysis and safety analysis is not considered to be placed under a unified framework.In response to the above problems,this paper proposes a vehicle trajectory learning framework for the intersection conflicts analysis.The framework is divided into several modules: the trajectory acquisition,the trajectory preprocessing,the path clustering and the application.Specifically,the innovative work of this article includes: First,a set of unsupervised vehicle trajectory learning frameworks for intersection conflict analysis was proposed and validated on multiple data sets;Secondly,the algorithm of scene ROI region clustering is proposed and verified on multiple datasets.The experimental results show that the algorithm can stably cluster the ROI region of the scene.Thirdly,a trajectory ensemble clustering strategy based on trajectory statistical features and distance measurement results is proposed,and the Laplacian centrality of the graph is introduced to find the representative of the cluster,and the redundancy clusters are merged according to the cluster representative.Fourthly,using the trajectory data and the previously extracted basic information of the scene,the trajectory arrival time prediction model is built.The safety and efficiency of the intersection are analyzed by the trained model.The methods proposed in this paper are validated on multiple data sets,and their effects and robustness are verified.