Research On Point Cloud Classification Method Based On Graph Convolution And Data Augmentation

Analyzing and processing point cloud data enables the real world to be better understood by computers.As an important branch of processing point cloud data,point cloud classification has crucial research significance and value in applications such as robotics,autonomous driving and smart cities.On the one hand,in view of the lack of topological information of point cloud data and its sparsity and irregularity,it is difficult for current point cloud classification networks to extract effective features;on the other hand,corruption of point cloud data is inevitable due to various factors such as the collection environment and methods,and existing training strategies lead to weak anti-corruption capabilities of point cloud classification models.Regarding the issue above,this thesis investigates the point cloud classification method based on deep learning technology.The research contents mainly includes:(1)Aiming at the problem that the universal methods of point cloud classification are difficult to extract effective features of point clouds,a siamese adaptive graph convolution algorithm(SAGCA)has been proposed.A siamese adaptive graph has been constructed to represent the topological relationship between point cloud data,and the original features have been weighted and fused by graph convolution to extract more critical features.In addition,in order to make the point cloud classification task can be better completed,SAGCA has been combined with various point cloud classification networks in both local and global ways,and two kinds of feature relation graphs also have been constructed in the siamese adaptive way.The local context information has been deeply mined by the former,and the overall topological relationship has been fully grasped by the latter,so that the feature extraction ability of the network has been enhanced.The performance of the algorithm proposed on the public datasets Model Net40 and Scan Object NN has been evaluated.The experimental results show that after SAGCA has been applied to different networks of point cloud classification,the classification performance of the networks is effectively improved,and it has certain advantages ove r the current mainstream classification methods.(2)Aiming at the problem of insufficient anti-corruption ability of present point cloud classification methods,a transformation-noise-density data augmentation method(TND)has been proposed.The point cloud training set has been subjected to morphological transformation,noise addition,and mixed density changes in sequence by this method,which has simulated common corruption what exists in reality,so as to the anti-corruption performance of the point cloud classification model has been significantly improved during the training process.The effectiveness of the proposed method has been verified on the public dataset Model Net40-C that contains various levels of corruption,The experimental results show that compared with other data augmentation methods,the robustness of multiple point cloud classification networks against corruption is significantly enhanced by TND.