Research On Pedestrian Object Detection Using Graph Neural Network And Laser Point Clouds

Recent years,the 3D sensing technology based on LIDAR for autonomous driving is in a booming stage of development.With the rapid development of deep learning and neural networks,the laser point cloud detection technology has also entered a rapid development stage.Among the existing detection algorithms,the detection accuracy of vehicle and cyclist is high,but laser point cloud pedestrian detection accuracy is low,and there are fewer pedestrianoriented research algorithms.According to the cutting-edge algorithms published in the KITTI dataset,the accuracy of vehicle detection reaches 90%,while the accuracy of pedestrian detection is only about 45%.Therefore,the main objective of this paper is to investigate a new method for laser point cloud pedestrian object detection with higher prediction efficiency.The basic idea is to improve the accuracy of 3D point cloud pedestrian detection by constructing a laser point cloud pedestrian skeletal dataset,building a laser point cloud pedestrian detection network,and establishing a graph structure based on pedestrian skeletal features.The main research of this paper is as follows.(1)Produce a laser point cloud pedestrian skeletal dataset 3DPL Dataset.Based on the KITTI open source pedestrian point cloud dataset,the bin files in the KITTI dataset are converted to pcd files,and then the pedestrian skeletal annotation is performed using Cloud Compare software to build a pedestrian skeletal point cloud dataset.By comparing 2D image and 3D point cloud map,annotate pedestrians,vehicles and cyclists in 3D point cloud map,select 11 key points for pedestrian graph construction,complete pedestrian skeleton point annotation in the point cloud map,and make 3D spatial coordinates of skeleton key points into txt files.(2)Build a two-stage laser point cloud pedestrian object detection network.In the first stage,the pedestrian skeleton generation network is designed to generate 3D bounding box proposals from bottom-up using Point RCNN network,and the model is trained using Bin-based method to generate bounding box proposals on the foreground points.Based on the point cloud within the proposal regions,Point CNN network is used to shrinkage point cloud,cluster pedestrian joint node and generate pedestrian skeleton.The second stage designs PP-GNN based on pedestrian refinement graph neural network,constructs graph relations using pedestrian joint aggregation points,generates skeletal point topology relations for pedestrian object detection,optimizes the size of the bounding box,and improves the detection accuracy.(3)Experimental validation.Detection validation is performed on the laser point cloud pedestrian skeleton dataset 3DPL Dataset and the official KITTI dataset,verifies the feasibility of the pedestrian point cloud detection network structure designed in this paper by comparing the accuracy and detection time with other detection algorithms.With the goal of improving the accuracy of laser point cloud pedestrian detection,this paper produces the laser point cloud pedestrian skeleton dataset 3DPL Dataset and designs a two-stage laser point cloud pedestrian detection network,generates pedestrian skeletons in the first stage and refines the bounding boxes in the second stage.Through experimental validation on the laser point cloud pedestrian skeleton dataset 3DPL Dataset and the official KITTI dataset,it is demonstrated that the detection network designed in this paper can effectively improve the accuracy of laser point cloud pedestrian detection.