Research And Implementation Of Roadside Scene Segmentation Based On Laser Point Cloud In Vehicle-road Collaboratio

Roadside perception plays a vital role in vehicle-infrastructure cooperation.Li DAR has the advantages of a large field of view and strong anti-interference ability.It is an essential sensor for roadside perception systems to obtain road scene information.The 3D data of road scene is collected by roadside Li DAR and semantic information is obtained by point cloud segmentation,which can provide essential data support for vehicle-infrastructure cooperation.Because the roadside Li DAR is deployed at a fixed location,the background area of the point cloud data acquired is usually unchanged.Around this feature,this thesis focuses on the research of point segmentation based on roadside Li DAR.The work is carried out from background extraction,ground point cloud segmentation,moving and static point cloud segmentation,and typical dynamic and stationary target point cloud classification in the foreground to provide rich perception information to the roadside perception system.Specifically,the main work and innovations of this thesis are as follows:(1)A roadside Li DAR point data collection scheme is designed.At the same time,it builds a collection system and uses two Li DARs to complete point cloud data collection for straight roads and crossroads.(2)Aiming at the fixed location deployment of Li DAR sensors in the roadside perception system,this thesis proposes a background extraction method based on the channel distribution of Li DAR.This method divides the sequence point cloud according to the Li DAR channel and the horizontal angle.Then the density cluster algorithm is used to select the qualified points as the background data.Finally,it introduces the scan line compensation method to amend the background information.Experiments with data from different Li DARs in multiple scenes prove this method can effectively extract the background information of roadside scenes.(3)A segmentation method of ground point cloud and moving-and-static point cloud in roadside scenes is proposed.It proposes a two-stage ground segmentation method for point cloud based on neighboring point features to obtain ground points in the background.For moving-and-static point cloud segmentation,firstly,it designs a method for generating the residual cloud image of the foreground in the roadside scene,introducing a three-dimensional semantic segmentation network to complete the moving-and-static points segmentation preliminarily.Finally,it proposes a Divide-and-Merge cluster algorithm using adaptive threshold-based to provide moving and static point cloud data with different cluster number information.The clustering algorithm adaptively changes the cluster threshold according to the characteristics of Li DAR data.This method can effectively improve clustering accuracy under the premise of satisfying real-time performance.(4)A point cloud classification method of typical dynamic and stationary target for foreground point cloud roadside scenes is designed.This method designs a way to extract key points of foreground point cloud clusters.It implements the data association of point cloud clusters between different frames based on the key points to distinguish the movement state of the point cloud clusters.It selects distribution,reflectance,and other features as the classification feature.It creates a point cloud feature data set based on the association results and trains a support vector machine classifier.The experimental results show that the classifier can effectively classify vehicles,pedestrians,and cyclists in the foreground point cloud.