Research On Semantic Segmentation Of Point Clouds Based On Graph Convolutional Neural Networks

Point cloud semantic segmentation has received increasing attention as the basis of 3D scene recognition and understanding,and has a wide range of applications in the fields of indoor navigation,robot localization and indoor layout.By performing point cloud semantic segmentation on 3D indoor scenes,it can help robots or navigation systems to accurately identify different objects and location distributions,such as walls,furniture,doors and windows,to achieve more accurate indoor positioning and layout.Traditional graphical convolutional neural networks ignore the geometric structure of objects when performing point cloud semantic segmentation,and fail to fully consider local features,resulting in the lack of clear object contours in the segmentation results and the overall segmentation accuracy cannot meet the actual needs.To solve the above problems,this study designs a graph convolutional neural network for semantic segmentation of point clouds in indoor scenes,aiming to fully extract point cloud features and provide higher accuracy and robustness for point cloud semantic segmentation tasks.The main research results are as follows:(1)The attention-based Graph Convolution Network(A-GCN)is constructed to address the problem that the traditional graph convolution model cannot capture the structural features of object contours in point clouds well.The A-GCN network is mainly designed with the Graph Attention Edge Convolution module(GAtt-Edge Conv)based on the attention mechanism.This module calculates the distance between the target node and other nodes by the K Nearest Neighbor(KNN)algorithm,sorts them according to the ascending rule,and selects the points corresponding to the top K values as the neighbor nodes of the target node.Edges are added between the target node and the K neighboring nodes to construct the undirected connectivity graph.On this basis,the attention weights of the neighbor nodes are determined according to the 3D coordinate differences and feature differences between the target node and the K neighbor nodes to distinguish the importance of the neighbor nodes in the undirected connectivity graph and extract the local structural features of the target node.The m Io U of the A-GCN network model on “area 5” of the S3 DIS dataset is 52.27%,which is 1.57% better than the base model DGCNN under the same experimental conditions.(2)To further improve the ability of local feature information extraction in A-GCN networks,a local feature aggregation algorithm,Net VLAD++,is designed in this paper.A Vector-based Graph Convolution Neural Network(V-GCN)model is constructed based on this algorithm.The model uses edge convolution and the Net VLAD++ algorithm to design a feature encoding module,which can effectively aggregate local features and different layers of features to improve the network model segmentation effect.The m Io U of the V-GCN network model on“area 5” of the S3 DIS dataset is 51.42%,which is 0.72% better than the base model DGCNN under the same experimental conditions.(3)To take full advantage of GAtt-Edge Conv and Net VLAD++ algorithms,we use both graph attention edge convolution features and local/global aggregation features To construct a graph convolutional network model with joint attention and vectors.(AV-GCN).The network can distinguish the importance of K-neighboring nodes of the target node and take into account the contour structure of the object,while effectively aggregating local and global features to significantly improve the feature extraction capability of point cloud.The m Io U of this model is 55.22% in “area 5” of S3 DIS dataset,which is 4.52% better than the base model DGCNN under the same experimental conditions.The m Io U in the “6-fold” cross-validation of the S3 DIS dataset is 61.65%,which is 1.61% better than the base model DGCNN.