Research On 3D Point Cloud Segmentation Algorithm Based On Graph Convolutional Neural Network

With the wide application of 3D scanners and depth sensors technologies,the approach to obtain 3D point cloud data has become more and more convenient,and the 3D modeling tasks based on deep learning have become a hot research topic.However,the 3D point cloud is disordered and irregular,and it is a kind of non-Euclidean domain data.It is still challenging to obtain its semantic features directly.In the idea of feature extraction,current deep neural networks such as PointNet network and PointNet++ network often combine the global features of 3D point clouds with local features at different scales,but ignore the structural information and position relationship between points.Graph,as a structured data type,can well express the relationship between nodes in a graph in the form of graph convolution.Therefore,the 3D point cloud segmentation methods based on graph convolutional neural network can be well applied to the modeling and processing tasks of discrete 3D point clouds.In order to enrich the semantic features of 3D point clouds,we propose two different graph convolution neural networks.The main contents include:(1)We propose an algorithm that fuses the direction and distance of points to better extract the local features of point clouds.We construct the similarity matrix by comparing the cosine similarity between points,and then use the k-Nearest Neighbor algorithm to select the most k points to form a local neighborhood graph.Next,we obtain edge features based on the distance between any two points in the graph,and use graph convolution to enrich the contextual semantic representation of point clouds.In addition,for the target classification task of the 3D point cloud,we combine the traditional cross-entropy loss function with the center loss function,thereby improving the classification accuracy.(2)We also propose another graph convolutional neural network algorithm to enhance the edge features of point clouds.We first search for the spatial neighbors of all points in a 3D point cloud and connect them into a global directed graph.Then we obtain the edge features of each edge in the graph by feature splicing of vertex position difference,node position offset feature and the position relationship between points.Next,we apply the attention mechanism to assign different weights to different edges in the graph to further strengthen the edge features.We also use Sobel operator to predict the semantic label of the 3D point cloud,which simplifies the algorithm complexity.In addition,in the process of the network training,the traditional crossentropy loss function is adaptively combined with the improved normal loss function to form the multivariate loss function,which can effectively improve the performance of 3D point cloud classification and segmentation.We have performed extensive experiments on the above two algorithms on object classification,part segmentation and indoor semantic segmentation tasks,involving three public datasets,namely ModelNet40 dataset,ShapeNet Part dataset and S3 DIS dataset.The experimental results show that the two algorithms show relatively mature effects on the classification and segmentation tasks of the 3D point cloud.Whether from the perspective of the Mean Intersection over Union(mIoU)of the semantic segmentation or the overall segmentation accuracy,both the two network models have a great improvement over the current mainstream 3D point cloud segmentation models.