Point Cloud Semantic Segmentation Based On Dynamic Graph Convolutional Networks And Feature Aggregation

With the rapid development of 3D point cloud data acquisition technology,the application of point cloud data is becoming increasingly widespread.How to achieve effective segmentation of massive point clouds to solve scene understanding problems is an important research direction in many fields.Point cloud semantic segmentation plays a fundamental role in the field of 3D scene understanding.However,due to the unordered and unstructured nature of point clouds,it is difficult to fully capture the local features of point clouds,which affects the segmentation accuracy.Therefore,this paper proposes two new point cloud semantic segmentation algorithms aimed at addressing two problems: how to learn features from the local structure of point clouds,and how to aggregate the local and global features of point clouds.The main contents and results include:Firstly,a dynamic graph convolution network,DGPoint,is proposed to enrich the local features of point clouds.This network introduces the Edge Conv module and dynamic graph convolution operation.Prior to Edge Conv,the K-nearest neighbor(KNN)algorithm is used to determine a new local area to achieve the effect of dynamic graph updating.The feature aggregation function in the Edge Conv module uses the dual-channel pooling(max and average)operation to compensate for information loss,which can better preserve the global and local features of point clouds.Although the DGPoint network achieves good segmentation results overall(m Io U of 68.3% and o Acc of 86.2% on S3DIS),repeatedly assembling local feature encoders to extract global features can cause semantic confusion in dense or similar-looking areas,which affects the overall segmentation accuracy.Then,aiming at the problems existing in the DGPoint network,a semantic segmentation network EPConv that aggregates local and global features of point clouds is designed.The network still uses the Edge Conv used in the DGPoint network to extract the local features of the point cloud,and adopts a hierarchical local feature extraction design to abstract global semantic features by aggregating local features.The sampling layer and the grouping layer use the farthest point sampling(FPS)and ball query grouping respectively,and the feature extraction layer uses the Point Conv module to solve the segmentation problem caused by the disorder and uneven density of point cloud data.In addition,in order to ensure the accuracy of edge extraction,the edge features extracted by each layer in the feature extraction layer are superimposed(add)and then input into the decoder for restoration.The decoder uses Point Deconv as a deconvolution operation to solve the feature propagation problem.The optimized network EPConv was experimentally evaluated in three different tasks,including part segmentation(Shape Net dataset),indoor scene segmentation(S3DIS and Scan Net V2 datasets),and road scene segmentation(Yellowstone dataset).The experimental results show that EPConv achieved excellent performance in each dataset,outperforming current state-of-the-art methods and demonstrating strong feature learning and generalization capabilities.In addition,through ablation studies,the effectiveness of each module in EPConv was further demonstrated.When both the Point Conv and Edge Conv modules were used in the network,local features of point clouds were effectively extracted and global features were aggregated,resulting in a more significant improvement in the segmentation results.