Research On 3D Target Detection Based On Deep Learning

With the advancement of science and technology,3D target detection has become a hot spot in the field of computer vision in recent years,and it plays a vital role in robotics and autonomous driving.How to make 3D target detection more accurate and rapid to identify objects has important practical application value and research significance.In order to effectively avoid the shortcomings of traditional 3D algorithms such as slow detection speed and low accuracy,and give full play to the advantages of deep learning in the field of image,this paper studies 3D object detection algorithm based on deep learning.In this paper,vehicles and pedestrians are the main research targets,and a 3D object detection algorithm based on deep learning is studied.The CARLA simulation platform is used to make Simulate point cloud data,and experiment on the three current excellent 3D detection algorithms PointRCNN,Pointpillars,and PV-RCNN.Based on the PV-RCNN network,CBAM attention mechanism,adaptive deformable convolution,and context are added.The fusion module and Gumbel Subset Sampling module make up for the shortcomings of difficult extraction of pedestrian features,low detection accuracy of distant objects,and the farthest point sampling algorithm can only sample from low-dimensional Euclidean space.The main work and research results of this article are as follows:First of all,in view of the high price of sensors that collect point cloud data in reality and the complicated point cloud data annotation,a point cloud simulation data platform was built using the CARLA simulation platform and a point cloud simulation data set was established.Then the automatic collection of sensor data and the automatic annotation of point cloud data in the simulation scene were realized by writing a script program;the validity of the simulated point cloud data was verified by comparing with the KITTI data set.Secondly,in order to find an excellent target detection network,the current three 3D target detection algorithms with good detection performance are experimentally compared.Conduct network training experiments on PointRCNN,PointPillas,and PV-RCNN respectively,and analyze the detection results of each network model.Experimental results show that the average 3D detection frame accuracy rates of PointRCNN,PointPillars and PV-RCNN for vehicle detection are 71.08%,68.61%,and 82.86%,and the average 3D detection frame accuracy rates for pedestrian detection are 82.94%,80.39%,and 80.87%.Finally,PV-RCNN was selected as the backbone network for the detection of simulated point cloud data sets.Thirdly,in view of the shortcomings of PV-RCNN network background information,noise and other interference information,the proportion of interference information is too large,the point cloud density distribution is uneven,the pedestrian feature is difficult to extract,and the farthest point sampling algorithm depends on the selection of the initial point.The network improved 3D target detection algorithm.Constructed CBAM attention mechanism,adaptive deformable convolution,context fusion module and Gumbel Subset Sampling module,and performed ablation experiments on the improved points.The experimental results show that the improved PV-RCNN network detection accuracy on the simulation point cloud dataset has an average 3D detection frame accuracy of 82.90%for vehicles,which is an increase of 0.27%over the previous improvement,and the average 3D detection frame accuracy for pedestrians is 83.36.%,which is an increase of 2.49%compared to before the improvement.Finally,the improved PV-RCNN detection model is verified on the original KITTI data set.The experimental results show that the improved network has a significant improvement in the detection of pedestrians and cyclists,and the average precision measurement accuracy has been increased by 2.68%and 4.87%,respectively,verifying the effectiveness of the improved PV-RCNN network.