Research On 3D Point Cloud Recognition Method Based On Hierarchical Graph Convolutional Neural Network

With the development of three-dimensional(3D)imaging technology,the acquisition of 3D data has become more and more convenient.As one form of 3D data representations,3D point cloud has unique advantages in reconstructing complex scenes and expressing objects shape.Therefore,utlizing 3D point cloud data for identification has gradually become a hot topic for applicational research.Indeed,this technology would advance the accuracy for perceiving the environment and identifying objects,such as agumented reality,intelligent manufacturing and autonomous driving.The present study,the researcher studies the 3D point cloud recognition methods based on the graph convolutional neural network.In particular,we investigated the point cloud classification algorithm based on the residual graph convolutional neural network and the point cloud semantic segmentation algorithm based on the hierarchical graph attention network.Our work realize the point cloud recognition scheme with high accuracy,low number of parameters and good portability,for the practical application needs such as 3D scene understanding and 3D target detection.The main innovations and achievements of this work are summarized below:(1)A point cloud classification algorithm based on the residual graph convolutional neural network is proposedIn order to address classification inaccuracy caused by the shallow layers that cannot extract the deep features of point clouds containing semantics,we propose a point cloud classification algorithm based on the residual graph convolutional neural network.By combining the design idea based on residual learning,the algorithm constructs a graph convolutional deep neural network to extract deep semantic features of point clouds,and it uses a graph pyramid pooling module with multi-scale feature fusion to further improve the recognition accuracy.It is experimentally verified that the classification accuracy of this algorithm reaches 93.4% on the synthetic 3D object dataset,which is 0.5% better than the shallow graph convolution model,and it also achieves 85.0% classification accuracy on the real 3D scanned object dataset.(2)A point cloud semantic segmentation algorithm based on the hierarchical graph attention network is proposedConsidering that the point cloud classification model focuses more on the recognition of a single object using global feature information,while the point cloud semantic segmentation task requires the recognition of multiple objects in the scene,directly extending the point cloud classification model to handle the point cloud semantic segmentation task will lead to a decrease in segmentation accuracy.To address this problem,a point cloud semantic segmentation algorithm based on the hierarchical graph attention networks is proposed.The algorithm takes full advantage of the ability of graph attention networks to better handle features at details,and recovers the spatial information lost due to the extraction of deep features at the encoder stage by applying the encoder-decoder hierarchical network structure.The experimental results show that this algorithm improves the segmentation accuracy by0.5% on the indoor 3D scene point cloud dataset and 0.7% on the fine-grained part segmentation point cloud dataset compared with the benchmark algorithm,and the visualization has more accurate detail segmentation results.(3)Implemented the porting and deployment of point cloud recognition algorithm on artificial intelligence chipWith the above point cloud recognition algorithm based on graph convolutional neural network as the core.Further,the trained model is deployed on an accelerator card based on an artificial intelligence chip after the steps of model conversion,model compression,and model acceleration.Experimentally,the deployed model can process the point cloud recognition task with less than one-tenth of the power of an NVIDIA graphics card which with an average decrease of 3.5% in accuracy.Therefore,our work implements a portable,low-power point cloud recognition module,and have significant implications for future research.