Research On 3D Point Cloud Target Detection Based On Spectral Voting Learning

At present,3D point cloud data has been widely used in the field of computer vision,and point cloud scene target detection using deep learning has become a research hotspot.Point cloud is a kind of unstructured data,and the existing deep learning network architecture is prone to problems such as inadequate feature extraction and poor generalization ability of the network framework.This topic takes spectral convolution as the core,integrates Hough voting and tensor voting theory,and carries out the research of 3D point cloud target detection based on spectral voting learning.The main research work of this paper includes:First of all,in order to solve the problems of insufficient feature extraction ability and low target detection accuracy in the existing 3D point cloud voting learning model,a SGCVNet point cloud target detection method based on spectral convolution is proposed.In order to solve the problem of insufficient feature information extraction,a spectral voting module based on spectral convolution and Hough voting theory is designed;By structurally combining spectral voting module,target candidate generation module and non-maximum suppression module,a SGCVNet point cloud target detection network is constructed,and experimental research is carried out on SUNRGB-D and Scan Net V2 data sets.Secondly,in order to solve the problem of insufficient utilization of global relations in the process of feature extraction by SGCVNet network model,a SA-SGCVNet point cloud target detection method based on self-attention mechanism is proposed.In order to make use of the global weight information between features,a self-attention model based on non-local position coding is constructed and integrated into the spectral voting module and the target candidate generation module;On this basis,a SA-SGCVNet point cloud target detection network is designed,and the experimental research is also carried out on two kinds of data sets.Finally,aiming at the poor robustness of voting model in SGCVNet network,a TV-SGCVNet point cloud target detection method based on controllable tensor voting is proposed.A voting layer based on controllable tensor voting theory is designed,and the controllable tensor voting layer is used as the voting process of the spectral voting module to improve the overall robustness of the module;Based on the above theory,the TV-SGCVNet point cloud target detection network is designed,and the experimental research is carried out separately on the ScanNetV2 data set.