Abstract: A density-oriented point cloud dynamic graph convolutional network is proposed to overcome the shortage in local feature extraction and ignoring the density of point clouds of current methods. Firstly, the point cloud local density index is generated to measure the density of points and their neighborhood points in the corresponding spatial location. Secondly, the local density index is used to dynamically give each point a dilated factor and group points as a local graph. The dynamic edge convolution extracts feature of each local graph. It not only extracts the geometric features, but also realizes the permutation invariance. Finally, the idea of residual network is used to optimize over-smoothing of graph convolutional network. The classification accuracy of the proposed network on ModelNet40 and ScanObjectNN is 93.5% and 82.2% respectively, and the mean IoU on the segmentation datasets ShapeNet and S3DIS are 85.6% and 60.4% respectively, which are higher than those of current networks such as DGCNN. The experimental results show that the accuracy in these tasks significantly improved, and it has good robustness in processing uneven density point clouds, which verifies the feasibility and effectiveness of the network.