Research On The Fundamental Algorithms Of Using Roadside LiDAR Sensors For Traffic Safety-Related Applications

Traffic safety is one of the major concerns for researchers and the public.Developing technologies provide more potentials for traffic safety research,and thus how to better improve traffic safety with new technologies becomes an important research topic.Connected/ Autonomous Vehicles(CAVs)technology is the cutting-edge technology in the field of intelligent transportation.This technology achieves all-around interaction and unification of humans,vehicles,roads,and the environment by integrating advanced communication technology,sensor technology,and big data algorithms,which has significant advantages and potentials in promoting the development of traffic safety.However,subject to the problem of carrying advanced sensors(ordinary vehicles and pedestrians are hardly equipped to carry advanced sensors),in the current stage,the popularization of CAVs technology still faces significant challenges,and its advancement in traffic safety is also severely limited.Light Detection and Ranging(LiDAR),a sophisticated sensor,is one of the advanced sensors used in the current CAVs technology environment(mainly in partial smart vehicles),which has the advantage of providing real-time,high-resolution traffic data without the influence of adverse lighting and most weather conditions.Unlike on-board LiDAR with a single service subject,The LiDAR deployed alongside highways,urban roads or inside intersections(roadside LiDAR)can provide data services for all pedestrians and vehicles within its detection range,which is considered as one of the ideal solutions to promote CAVs technology.Considering that the related research based on roadside LiDAR is still in a development stage,and there is still much room for improvement in roadside LiDAR data processing algorithms to support traffic safety-related applications,this paper conducted in-depth studies on data processing algorithms,including point cloud data pre-processing,object velocity and trajectory tracking,and automatic lane identification.Eventually,a series of improved fundamental algorithms were proposed.Notably,to highlight the supporting role of the proposed algorithms the proposed algorithms were directly applied to near crash recognition,a significant traffic safety-related application.According to the research contents,the main highlights of this paper are summarized as follows:(1)The roadside LiDAR data pre-processing algorithms were optimized,involving background filtering,object clustering and object classification.A BP neural network-based background filtering algorithm with both stability and resistance to pseudo-dynamic/static point interference was proposed to address the problems of the background filtering.Tests show that the precision rate and recall rate of the filtering algorithm can reach more than 95%.The modified algorithm based on hierarchical density clustering(HDBSCAN)is proposed for the clustering problem of LiDAR point cloud with uneven density.The test results show that the modified HDBSCAN has the expected stable clustering effects.To solve the problem of the accuracy of pedestrian/vehicle classification from point cloud clusters under severe occlusion,a classification feature with the ability to resist occlusion interference,regular straight edge(RSE),was proposed,and the corresponding RSE extraction algorithm was designed and validated.The tests show that the classification performance in BP-ANN and Random Forest classifiers is improved by around 2% since using RSE for training.(2)A convolutional multi-point matching velocity estimation algorithm was proposed for the vehicle speed tracking problem based on roadside LiDAR data.The algorithm was built with a multilayer structure,involving object localization,point cloud imagery,image matching,matching point extraction and velocity filtering,aiming to solve the defect of unstable accuracy for velocity estimation with fixed reference points.In addition to constructing an estimation strategy based on multiple matching points from image matching,the algorithm also contributes an image normalization algorithm designed to improve the matching accuracy,and a matching correction-based Gaussian filter algorithm for enhancing the estimation stability.The tests show that convolutional multi-point matching velocity estimation algorithm has higher accuracy and stability compared to the poplar fixed reference point method.(3)For the vehicle trajectory tracking problem based on roadside LiDAR data,an adaptive Kalman filter-based vehicle trajectory optimization algorithm was proposed.The algorithm designed a noise covariance adaptive correction strategy by combining the stable estimated velocity(provided by the previously proposed velocity estimation algorithm)and the characteristics of roadside LiDAR data,aiming to address the problem of the traditional Kalman filter whose performance is seriously degraded due to the incompatible noise prior assumption.The numerous cases show that the adaptive Kalman filter-based vehicle trajectory optimization algorithm has significant improvement in the smoothness and reliability of the filtered trajectory compared to the traditional Kalman filter.Moreover,the trajectory optimization algorithm can be used under simple assumptions,which is very friendly to users.(4)An improved automatic lane identification algorithm(IALIA)was proposed for automatic lane identification using roadside LiDAR data.IALIA is a three-stage algorithm covering RES trajectory extraction,basic lane contouring,and lane segmentation based on pedestrians crossing the street(the third stage is optional).Compared with existing roadside LiDAR-based automatic lane identification algorithm,IALIA has the advantage of higher automation and wider applicability(applicable to lane identification at intersections).Finally,the intersection example validated the effectiveness and automatic lane identification performance of the IALIA algorithm,indicating the algorithm’s potential for future practical applications.The research results of this paper can provide algorithm support and technical reference for follow-up roadside LiDAR-based research(not limited to traffic safety-related research),as well as have great benefits for promoting the practical application of the roadside LiDAR.