Roadside LIDAR-based Traffic Target Sensing Method And Implementation

Accurate sensing of traffic targets plays a crucial role in improving traffic supervision and optimizing the efficiency of intelligent transportation.With the advantages of strong anti-interference capability and high spatial resolution,LIDAR has been deployed as the main sensor for intelligent transportation and vehicle-road cooperation systems in recent years,and is used for real-time traffic point cloud data collection.Due to the existence of redundant background point clouds and inter-target occlusion in real road scenes,as well as the characteristics of near-dense and far-lost point cloud data,existing target sensing methods still have problems such as lack of accuracy and large computational effort.Therefore,this thesis proposes a roadside LIDAR-based traffic target sensing method,and carries out research in four main aspects,namely ground plane segmentation,target point clustering,traffic target identification and tracking,as follows:(1)In the process of ground plane segmentation,to address the problems that the existing methods cannot accurately fit the actual ground plane with slope changes,we firstly propose an improved sector grid model based on cylindrical coordinate system and the lowest point representative(LPR)method to optimize the selection of seed points required to fit the plane,then design a multi-ground plane model and combine the random sampling consistency algorithm(RANSAC)to fit the ground points in each region separately.Finally,the continuity of each fitted plane is determined to achieve the complete segmentation of the ground plane.The experimental results show that the segmentation accuracy of the ground point cloud in four typical road scenarios is more than 86%,which can achieve the accurate segmentation of ground plane in different road scenarios.(2)In the target point clustering task,an improved Euclidean clustering algorithm is proposed for the problems that it is difficult to adapt a single fixed distance threshold to the targets at different distances and to set the distance threshold repeatedly manually based on empirical values.The algorithm firstly accelerates the clustering process by constructing a three-dimensional KDTree data structure,secondly divides the space in which the point cloud is located into multiple clustering regions by the distribution law,and finally performs the adaptive calculation of distance thresholds for each clustering region to complete the accurate clustering and detection of non-ground point targets at different distances.The experimental results show that the algorithm is highly efficient and the accuracy of the improved algorithm for target clustering at different distances is significantly improved compared with other clustering algorithms.(3)In the traffic target recognition,to address the problems that it is difficult to accurately distinguish different traffic target categories and optimize the classifier selection with a single feature in the sparse point cloud environment,firstly,the target spatial structure features,the longitudinal contour height sequence features and the overall distribution features of the point cloud are selected and constructed into a multi-dimensional composite feature input vector dedicated to the traffic targets in this thesis.Then,SVM is used as the base classifier and iterative reinforcement by Ada Boost algorithm,and finally the classification and recognition of traffic targets are realized.After experimental verification,the average accuracy of the method recognition is 89.2%,which is improved by 7.3% compared with the standard SVM algorithm,and the F1 value of each category is over 83%,which can meet the accuracy requirement of target recognition.(4)In traffic target tracking,an improved nearest neighbor tracking algorithm is designed to address the problems of partial occlusion between targets or changing point cloud patterns due to position changes,which leads to target tracking loss.Firstly,we construct the disparity equation by fusing intensity features,then calculate the disparity value between neighboring frames and generate the nearest neighbor association matrix,and finally associate the targets with non-zero minimum values to complete the tracking of traffic targets and the acquisition of information such as target speed and trajectory.The experimental results show that the tracking accuracy of the algorithm reaches 89.6% in the range of 60 m,which basically meets the accuracy requirements of the tracking task.In this article,the proposed method is experimentally validated by building a roadside point cloud data acquisition platform and combining DAIR-V2 X roadside data set,which improves the accuracy of the algorithm of each task module and optimizes the computational efficiency of the algorithm while ensuring the practicality of the proposed method,showing the good effect of the method in this thesis.It has certain practical application value.