Research On Discrete 3D Point Cloud Processing Based On Convolution Neural Network

In recent years,with the continuous development and maturity of depth sensor technologies such as lidar equipment and RGB-D cameras,the quality,acquisition efficiency and cost-effectiveness of 3D point cloud data have been continuously improved,and they have been widely used in unmanned driving,smart cities,surveying and remote sensing,etc.Fields,industries.As the main data format of 3D reconstruction technology,point cloud data has been widely used in classification and segmentation tasks of 3D models and scenes.Among them,the 3D point cloud classification and segmentation method based on deep learning has attracted more and more attention and research due to its superior performance.Point Net and Point Net++,as the earliest point cloud processing series network,solved the problems of point cloud disorder,but still faced the problem of insufficient local feature extraction.Based on the Point Net++ network,this paper introduces the self-attention mechanism into Point Net++ to construct the point cloud semantic segmentation network SSA-Point Net++ to further improve the accuracy of point cloud semantic segmentation in complex scenes.The main work of this paper is as follows:1.Firstly,this paper summarizes the current research status of 3D point cloud segmentation based on deep learning models from the three aspects of multi-view,voxel,and discrete point cloud;Then it briefly reviews the background knowledge and development of convolutional neural networks,and explains the theoretical basis of attention mechanism used to improve the network.;Finally,the public data set and evaluation index used in the 3D point cloud segmentation deep learning of this article are explained.2.The classic direct point cloud semantic segmentation neural network models Point Net and Point Net++ are verified on the Model Net40 and S3 DIS data sets,and the advantages and disadvantages of the network are analyzed,which lays the foundation for the subsequent network model design and performance improvement.3.Based on the Point Net++ network,this paper introduces the self-attention mechanism into Point Net++ to construct the point cloud semantic segmentation network SSAPoint Net++.And the spatial self-attention mechanism is clearly divided into sampling point and neighborhood self-attention mechanism,combining the two and adopting different spatial encoding methods to enhance the learning ability of the sampling point neighborhood topology;Build an attention pooling module and integrate multiple global features extracted by the attention mechanism pooling and maximum pooling through the differential pooling function,and use the spatial self-attention mechanism to distinguish Self-adaptive screening of sexual characteristics to strengthen the transmission of important information on the network.Experiments on public data sets Model Net40,S3 DIS and Semantic3 D show that the proposed network outperforms the existing networks in terms of both the overall accuracy and m Io U.The classification accuracy on the 3D model data set Moedl Net40 reaches 93.1%,which is 1.2% higher than that of Point Net++.For indoor dataset S3 DIS,the m Io U of the proposed SSA-Point Net++ is the best and 6.3% higher than Point Net++.While for outdoor dataset Semantic3 D,the m Io U of the proposed SSA-Point Net++ is about 3% higher than the sub-optimal.Compared with the segmentation results of other networks on the public dataset,proposed network is more general on different datasets and high application potential.