Research And Implementation Of Real-Time Road Obstacle Target Detection Algorithm Based On LiDAR

With the rapid development of laser detection and depth sensing technology,Li DAR has received more and more attention from researchers,including autonomous vehicles,autonomous navigation robots,and elevation remote sensing technology.In the field of autonomous driving,Li DAR detects the three-dimensional spatial information of the surrounding scene,enhances the spatial understanding of the surrounding environment by the autonomous driving system,and provides an effective basis for sensing downstream task such as path planning and behavioral decision-making.In the perception task in the field of autonomous driving,with the increasing maturity of sensing technology,the spatial information and target object characteristics contained in the 3D point cloud obtained by Li DAR are becoming more and more abundant.make great progress.This paper mainly studies the obstacle perception algorithm based on Li DAR in autonomous driving scenarios.The specific research contents are as follows:(1)Development of obstacle point cloud segmentation and clustering algorithms.The classical random sampling consistency algorithm(RANSAC)is inefficient in processing large point clouds of autonomous driving scenes.This paper analyzes the reasons for this phenomenon and designs an efficient ground point cloud segmentation algorithm.The direction random sampling consistency(O-RANSAC)algorithm,the calculation amount and parameter amount of this algorithm are reduced by about 30% on the basis of the original RANSAC algorithm,which improves the efficiency of processing large point clouds.This paper analyzes the advantages and limitations of common 3D point cloud clustering algorithms,and proposes an adaptive depth hierarchical clustering algorithm in which the clustering conditions change with the change of depth information,which improves the clustering effect of obstacles in different depth ranges.The problem of missing long-distance obstacle information of the laser sensor is avoided.Aiming at the phenomenon that the bounding box of the parallel coordinate axis lacks the target orientation information,this paper generates a foreground bounding box containing orientation information for each independent obstacle based on the label method.(2)Development of classification neural network based on point cloud clustering.This paper proposes the Point GLOH＿cls target classification network.The directional convolution of the network uses the semantic connection between the neighboring point clouds to convert the discrete and disordered independent point clouds into a compact and accurate vector representation,so as to make the correct point cloud target category prediction,which makes up for the shortcoming of missing part of the context information of the point cloud in the model based on the symmetric undirected calculation to extract the point cloud feature.In the target recognition tasks of the public datasets Model Net40 and Shape Net Core,the Point GLOH＿cls network has achieved SOTA-level performance.Point GLOH＿cls takes the clustered point cloud data in the bounding box in(1)as input,performs the target recognition function,and outputs the category probability vector of the point cloud cluster,which realizes the detection of scene obstacles by the Li DAR algorithm.(3)End-to-end point cloud target detection neural network development.The Point GLOH＿sem target detection network proposed in this paper uses the Point GLOH＿cls network as a point cloud feature extractor,directly acts on the 3D point cloud of the autonomous driving scene,and generates obstacle bounding box proposals based on the point cloud foreground segmentation;and then combines global features and local features to adjust the 3D point cloud.The parameters of the proposal,output the pose and size information of obstacles,generate accurate 3D bounding boxes for foreground points,and realize the end-to-end point cloud object detection function from data to results.(4)Test verification.The algorithm and network model proposed in this paper are verified on the test platform.The results show that the algorithm process and deep model proposed in this paper can effectively detect scene obstacles and provide a reliable basis for downstream tasks.