| With the increase of motor vehicle ownership,urban traffic safety has become an important factor that restricts the social and economic development of China and affects the quality of life of residents.Most urban traffic accidents occur because drivers or pedestrians do not have enough perception of the surrounding traffic environment.To enable pedestrians and motor vehicles to avoid traffic risks in advance,it is necessary to acquire the trajectory information of traffic targets in real time through sensors and predicate their behavioral intentions.In recent years,3D Li DAR has made good progress in target detection and tracking.Using roadside Li DAR as a sensor,this paper proposes background filtering,target identification,trajectory tracking and intention prediction methods for large traffic volumes and complex road environments.The main contributions of this paper are as follows:(1)This paper proposes a background filtering method based on the change of point cloud density,which improves the accuracy of the global background filtering of the low-channel Li DAR.Firstly,the roadside detection scene is spatially segmented,and the spatio-temporal distribution characteristics of Li DAR point cloud are analyzed.Secondly,the density change operator is constructed to extract the traffic participant passing area,and the noise of the traffic participant passing area is removed by DBSCAN algorithm.Finally,the background point cloud is removed based on the constructed background.Experimental results show that the background filtering accuracy of the algorithm can reach 99.81% and meet the real-time requirement of data processing.(2)This paper proposes a point cloud feature-based pedestrian-vehicle classification algorithm combined with machine learning methods.Firstly,the current frame point cloud is clustered,and each traffic participant is clustered into a single target.Secondly,a total of 24-dimensional classification features including detection target space volume description,geometric shape description,target point cloud quantity,target distance from detector,maximum echo intensity of target point cloud,and frequency distribution of point cloud echo intensity are constructed.Finally,the random forest method is used for pedestrian-vehicle classification based on features.The results show that the precision and recall rate of the classifier are 98.90% and99.45%,respectively,which can accurately identify pedestrian and motor vehicle targets from close to far distances.(3)This paper proposes a high-precision trajectory tracking and intention prediction method considering occlusion conditions to address the tracking errors caused by vehicle occlusion in high-density traffic scenarios.Firstly,a trajectory tracking method based on fixed tracking points is proposed for the stable tracking of vehicles in high-density traffic flows.Secondly,the target occlusion identification model is proposed based on the spatial position and point cloud completeness of vehicles,then the Kalman filter is used to repair interrupted trajectories of pedestrians and vehicles.Lastly,a long-short-term memory artificial neural network is employed to predict the intentions of traffic participants.The proposed trajectory tracking method achieved an average accuracy of 98.26% for vehicles and 89.91% for pedestrians.The trajectory prediction errors were 0.3606 m and 0.1675 m,respectively,and the method can connect interrupted trajectories effectively,and improve the accuracy of pedestrian and vehicle intention prediction. |
