Research On 3D Point Cloud Target Detection Method

In recent years,the rapid development of deep neural network based on deep learning makes great progress in the field of target detection,especially in the field of 2D target detection.However,2D object detection only returns the pixel coordinates of the target object and lacks depth information,which makes it has certain limitations in practical application.Therefore,the target detection technology based on 3D point cloud has become a research hotspot in the field of computer vision.After summarizing and studying the research status at home and abroad,two target detection algorithms based on 3D point cloud for outdoor and indoor scenes are proposed in this paper respectively.The feasibility and effectiveness of the proposed algorithm are verified by a large number of comparative experiments with existing3 D target detection algorithms.Firstly,the basic knowledge of 3D point cloud target detection technology is introduced in this paper,including the basic concept of point cloud,point cloud target detection,computer vision attention mechanism,Hough transform and depth Hough voting,data set and its evaluation index.For outdoor scenes,a 3D point cloud target detection algorithm based on visual attention mechanism is proposed in this paper.The 3D features and 2D features corresponding to input point cloud and point cloud height map are learned and fused through feature extraction network.Then,the position of the new object in the scene is focused on by the target localization network in each iteration,and the 3D object attention area is selected in the scene.Then 3D Transform and resampling module are introduced to solve the differentiability and speediness of data processing.The performance of 3D target detection is further improved by the classification and regression network from coarse to fine.Finally,the algorithm is packaged into software based on QT,which is convenient for users to operate.For indoor scenes,a 3D point cloud target detection algorithm based on 2D-3D voting is proposed in this paper.The input of the algorithm consists of two parts: RGB image and point cloud data.Firstly,the seed point features of the input point cloud data are obtained through the improved Vote Net,and then the 2D voting features obtained by the 2D image voting module are fused with the seed point features,and the target detection performance of the algorithm is improved by adding RGB image features.Then feature fusion and multi tower training module are introduced to make full use of the features from all modules.Finally,the algorithm is packaged into software based on QT,which is convenient for users to operate.Finally,pytorch1.3.1 is used to train,verify and test the model under ubuntu18.04 in this paper.The experimental results show that the proposed algorithm has better performance in3 D target detection.