Research On Voxelized Point Cloud Recognition Of Indoor Scene Based On Graph Neural Network

As one of the crucial technologies of semantic mapping,scene recognition technology can provide prior knowledge for service robots and improve the efficiency of robot operations.The scene recognition algorithm based on 3D point cloud which could make full use of the geometric information of the place and is almost not affected by illumination,has broad application prospects.With the development of deep learning theory,in recent years,the method based on graph neural network has become a popular research direction in the field of 3D point cloud recognition.The point cloud processing algorithm based on graph neural network can be roughly divided into point-wise processing method and voxelization method.The former operation is straightforward and simple,but to calculate the relationship between every two points,the computational complexity is too high to adapt to large-scale point clouds.The latter processes point clouds in voxels,which can adapt to largescale point clouds,but it is unavoidable to use 0 elements to pad,which increases the amount of calculation,and the added 0 elements will interfere with the task target,and the spatial relationship is also not clearly expressed.Furthermore,the classification accuracy of the two methods is almost only verified on the object classification dataset with small point cloud scale and uniform distribution,and has not been evaluated on the point cloud scene classification dataset.Aiming at the service robot application,this paper proposes a voxelized indoor point cloud scene recognition method based on graph neural network.The main research contents are as follows:(1)An end-to-end voxelized point cloud classification method is proposed,which is suitable for real point cloud scene classification tasks.The algorithm uses a graph convolution network to obtain discriminative voxel features for each voxel,and then uses the transformer attention structure to learn the relationship between voxels to complete the scene classification task.(2)Optimized for the shortcomings of traditional voxelized point cloud processing algorithms.By removing invalid voxels and downsampling within the voxels,the padding of 0 elements is reduced,and the computational complexity is reduced.In the voxel feature extraction and voxel feature integration,the initial spatial structure information is used to alleviate the interference of 0 elements to the effective information.Combine the knowledge of geometric descriptors with deep learning to enhance the discrimination of voxel features and improve the accuracy of scene classification.(3)The algorithm proposed in this paper is verified on three existing scene point cloud databases,and compared with the advanced graph neural network algorithms PointNet and DGCNN.The results show that the classification accuracy of the algorithm in this paper exceeds the advanced graph network algorithm DGCNN algorithm by more than 2%on a variety of real scene databases,and the highest is more than 3.8%,far exceeding the PointNet algorithm.A variety of ablation experiments are designed to analyze the rationality of network related structure and key parameters.Based on the experimental results,the rationality and inadequacy of the proposed algorithm in the scene point cloud classification task are analyzed.(4)The proposed method was transplanted onto robot mobile platform to evaluate the distinguishing abiliby of continuous frame point cloud scene.A frame fusiong method was introduced to enhance the ability of proposed algorithm on continuous frame scene which was capture from robot perspective.This frame fusiong method strenghs the utility of proposed algorithm in robotics applications.