Semantic Segmentation Of Point Clouds Based On Graph Convolutional Networks

3D Point Cloud Semantic Segmentation as a key technology in computer vision fields such as target detection,3D reconstruction,and scene understanding,is widely used in human intelligent life,and it has also become a hot and difficult point in the field of 3D vision re-search.At present,the research on 3D point cloud semantic segmentation tasks is mainly concentrated in the field of deep learning.However,there are two problems in these models:insufficient local feature expression and long training period.Aiming at these major limita-tions,this paper combines the advantages of traditional segmentation algorithms and graph convolutional networks,and proposes a 3D point cloud semantic segmentation model based on geometric segmentation and graph convolutional network,which can simultaneously im-prove the point cloud semantic segmentation accuracy and the efficiency of model training.Furthermore,in order to fully express the context of adjacent nodes,graph attention mech-anism is introduced on this basis,a 3D point cloud semantic segmentation model based on graph attention convolution is proposed.The specific work of this paper is as follows:(1)A 3D point cloud semantic segmentation model based on geometric segmentation and graph convolutional network is proposed.The model consists of four parts: geometric seg-mentation module,graph construction module,Point Net feature embedding module,and contextual semantic segmentation module.As the core part of the geometric segmentation module,the purpose is to sparse and express the original point cloud.This paper fully com-bines point features,point cloud local features and global features,redesigns the feature descriptor,and divides the points based on the feature descriptor and energy segmentation algorithm.The cloud is divided into ”superpoint” with a certain shape.Subsequently,the graph topology structure is established between the ”superpoint”.And the Point Net feature embedding module is used to perform feature embedding and alignment.Finally,the con-textual semantic segmentation module is used for semantic segmentation,and it is traced back to the original point cloud to obtain the semantic segmentation label.For the S3 DIS,the mIoU and the oAcc reach 58.6% and 86.4%,and the training period is greatly reduced.For the Semantic3D,the mIoU and oAcc reach 73.9% and 94.0% respectively.(2)A 3D point cloud semantic segmentation model based on graph attention convolutional network is proposed.The 3D point cloud semantic segmentation model based on geometric segmentation and graph convolutional network does not consider the influence of adjacent nodes of the graph on each other,so the model cannot make correct segmentation between similar objects or overlapping areas of objects.In response to this shortcoming,this paper introduces the graph attention mechanism based on the graph convolutional network,ex-plores the adjacent relationship between superpoints,assigns different weight coefficients to different neighboring points,and aggregates local information to obtain its local feature.After graph pooling and up-sampling interpolation,the final semantic segmentation result is obtained.Experimental results show that the model can obtain more contextual information from neighboring nodes,making the semantic segmentation result closer to the ground truth.For the S3 DIS,on the basis of the previous model,the mIoU and oAcc increases by 3.2%and 1.1% respectively.For the Semantic3 D,the mIoU of this model increases 0.7% and the oAcc decreases 1.1%.