Research On Traffic Multi-object Tracking And Information Extraction Techniques Based On Roadside LiDAR

In the context of building National Strength in Transportation and digital transportation,innovation and data empowerment have become the development direction for upgrading and improving the quality of the future transportation industry.The transportation industry has been continuously promoting the deep integration of advanced information technology with the traditional transportation industry.However,during the process of integration and development,the current limitations of intelligent transportation applications become apparent,particularly in terms of the absence of high-resolution,lane-level,and trajectorylevel data.These limitations hinder the ability of intelligent transportation applications to conduct in-depth analysis of traffic issues,accurate evaluation of traffic behaviours,and precise traffic safety control.Traditional roadside unit detectors,such as loop detectors,magnetic induction,and microwave radar,can only reflect the traffic flow conditions in the detection section and cannot recognize and track the trajectory of individual traffic flows.Although video detection can obtain the movement trajectory of vehicles in a relatively small local area,high-definition cameras must be installed according to the lane arrangement,and this sensor is seriously affected by light and shadow and cannot obtain reliable depth information.Currently,various detectors mainly present traffic data information with statistical features and have certain macro characteristics,but there are obvious shortcomings in terms of coverage,resolution,and information volume,which cannot meet the needs of road traffic flow deep analysis for smart traffic innovation applications.To meet the needs of highresolution,high-precision,and regional traffic flow data under precise traffic control,this paper uses roadside Li DAR to monitor roads and collect traffic data,achieving highresolution micro traffic information extraction,as detailed in the following research:(1)To address the current problem of low run-time performance and loss of valuable background points in roadside Li DAR filtering algorithms,this paper innovatively proposed the concept of point cloud slice unit and a filtering and segmentation method based on the slice.Firstly,based on the slice unit,the points in the point cloud are finely divided into four categories: valuable object points,worthless object points,abnormal ground points,and normal ground points.By using the projection filtering algorithm based on slice,the original point cloud is filtered at the fastest speed compared to similar algorithms,and in addition to traffic object points,the valuable background points that can help analyse traffic object behaviours can be preserved.Then,based on the filtered results,a region growth method was proposed to achieve instance segmentation of the point cloud.The attributes of slice are utilized to optimize the segmentation results for occlusion scenes,further improving the segmentation accuracy.In addition,a machine learning model based on slice features was proposed to classify various objects in the point cloud,extracting semantic information from the point cloud,demonstrating the high potential of slice units in point cloud processing issues.(2)To address the common issue of inadequate detection range in current traffic object detection based on low-channel roadside Li DAR,a traffic object detection range extending method based on static background construction was proposed.Firstly,continuous point cloud frame data covering the entire horizontal and vertical angles of the sensor is used,and the full-angle background points are obtained based on the distance difference between background objects and traffic objects,constructing a static background point cloud based on distance information.Furthermore,a static background optimization method was proposed to reduce the impact of traffic objects on background construction in heavy traffic scenes,reducing noise points in the background point cloud and supplementing missing far-distance background points.Compared to other background construction methods,it can construct the most accurate background point cloud with the least data.Using the difference between the static background and target point cloud to achieve precise point cloud filtering,extracting motion features based on the objects’ movement direction and distance,combined with the fast Fourier transform algorithm to filter noise points and obtain distant traffic object points.Experiments show that the proposed method is superior to other similar methods in both detection accuracy and range.(3)This paper proposed a lane-level map generation method based on low-channel roadside Li DAR to address the current challenges of generating lane-level maps that are not timely and distributed generated and updated.For the intersection lane-level map,the paper first presents a mathematical model to describe the signal phase,geometric shape,layout,and lane direction in the intersection lane-level map.Then,based on the result of traffic object detection,group trajectory is used to obtain the signal phase information of the intersection.Next,a convex hull-based detection algorithm is used to recognize the layout and geometric shape of the intersection.Subsequently,a sliding window algorithm was proposed to detect the lane markings and classify the lanes.Finally,a trajectory matching method was proposed to generate the connectivity of the turning lanes.For the road section lane-level map,the sliding window algorithm applied to the intersection is improved to be applicable to the lane marking detection in road section scenes,and the direction of the lanes in the road section is determined based on the driving patterns of vehicles.The experiments conducted on different types of intersections and road sections demonstrated that the proposed method meets the lane-level accuracy requirements and has advantages of lightweight,low-cost,and timely compared to the lane-level map generation methods based on satellite,mobile mapping system,and crowdsourcing.(4)A multi-vehicle tracking method assisted by self-generated lane-level maps was proposed to address the problem of obtaining complete and full-life cycle trajectories of vehicles in scenarios with low appearance information and frequent and long occlusions.Firstly,all the possible historical trajectories associated with the target vehicle are found through a search process constrained by the lane-level map.Then,the correct historical trajectories that can be associated are determined through a micro-motion model constrained by the lane-level map.Additionally,a detection optimization method based on the lane-level map was proposed to further improve the tracking performance.Finally,to facilitate a more comprehensive evaluation of the multi-vehicle tracking method based on roadside Li DAR,a tracking dataset suitable for roadside Li DAR is collected,annotated,and open-sourced.Compared with similar multi-object tracking algorithms,the proposed tracking algorithm has the highest accuracy and the strongest anti-occlusion ability and is more suitable for obtaining complete trajectories of vehicles within the scanning range of low-channel roadside Li DAR.(5)Based on the results of multi-vehicle tracking,high-resolution microscopic traffic information is extracted from high-resolution trajectory data.In terms of extracting microscopic traffic parameters,a Kalman-based vehicle centre position estimation model was proposed,which overcomes the impact of occlusion on the accuracy of motion parameter extraction and can represent occluded vehicle positions more accurately compared to the midpoint-based position estimation method.In terms of microscopic following models,the wavelet transform filter was used to optimize the preceding vehicle trajectory data,improving the following accuracy of classic following models such as Newell,OVM,IDM,etc.in the scenario based on roadside Li DAR.In terms of microscopic lane-change models,a Transformer-based lane-change intention and trajectory prediction model were proposed,and it is verified that the Transformer model has advantages over RNN and LSTM models in processing time-series data of microscopic traffic.