Research And Application Of 3D Point Cloud Target Detection And Tracking Technology Based On Deep Learning

Target detection technology and target tracking technology are very important research directions in the field of computer vision,It is widely used in the fields of autonomous driving,autonomous navigation,robot and virtual reality With the popularization and development of 3D point cloud acquisition equipment,3D point cloud data has brought a novel solution for target detection and tracking technology.In this context,this paper focuses on the research of target detection and tracking technology of 3D point cloud,and proposes an improved 3D point cloud target detection network and multi-target tracking framework based on SECOND network and 3d-mobile netv2 network.The main work and achievements are as follows(1)Based on SECOND network,this paper proposes a point cloud target detection network with higher accuracy and stronger real-time performance--SECOND++.This paper optimizes the three-stage feature generation network of SECOND network,and designs res2 senet network by combining multi-scale feature receptive field and attention mechanism,which can reduce the loss of spatial location information.In addition,in order to further improve the detection accuracy of complex point cloud target and make the classification confidence consistent with the IOU,this paper improves the convolution model and focus loss to mine semantic information more accurately and improve the detection effect.The experimental results show that the performance of SECOND++ is better than that of SECOND,avod-fpn,f-pointnet,voxelnet and mv3d.The running time is 0.025 seconds,the average detection accuracy of 3D target frame is 79.86%,and the average detection accuracy of aerial view is 85.18%.(2)In this paper,a multi-target real-time tracking framework based on 3d-mobile netv2 is proposed.The framework uses neural network to predict the state of the target object in three-dimensional space,employs the Hungarian algorithm for data association frame by frame,designs trajectory management module to manage the corresponding trajectory,and realizes multi-target tracking.Compared with the traditional framework,this framework does not need to perform Kalman filtering in the image space.It can not only achieve good performance at high frame rate,but also has excellent performance for occluded targets.Tested on Kitti dataset,the comprehensive performance is better than that of Kalman,MDP,lp-ssvm,DSM and complexer Yolo mainstream framework,The main indicators are as follows: FPS is 39,Mota is 79.22,motp is 78.33,MT is 54.19,ML is 13.12,IDS is only 16.