Research On Recognition And Segmentation Method For 3D Point Clouds Based On Graph Neural Network

In the past few years,3D deep learning has become an important research topic.Many application scenarios need to perceive the surrounding 3D environment,and sometimes need to interact with 3D objects.Point cloud has increasingly become a more popular data form in the field of 3D deep learning.In view of the characteristics of disorder and irregularity of 3D point cloud data,how to effectively use the spatial distribution information of objects or scenes,extract internal geometric information,and mine potential feature information has become the focus of research in the field of 3D point clouds.Main researches in this thesis are summarized as follows:(1)The deep learning methods based on point cloud data are summarized and analyzed: the three-dimensional recognition and segmentation methods of related point cloud data are analyzed,their respective advantages and disadvantages are summarized,and related datasets are introduced.(2)A three-dimensional point cloud recognition and segmentation method based on graph convolution network is proposed: a neighborhood graph is constructed in the neighborhood area of each center point to reflect the spatial distribution information of the neighborhood area,and then the graph convolution is used to extract potential geometric information.We considered the selection of neighborhood points in two spatial dimensions,so that not only the geometric distribution of the neighborhood area is fully mined,but also the information of similar feature areas in the feature space dimension is extracted.We fuse these two spatial dimension features to obtain a richer feature representation.Compared with some existing methods,this method has certain competitiveness.(3)A three-dimensional point cloud recognition and segmentation method based on graph attention network is proposed: the encoder-decoder network structure is used,and graph attention convolution and graph attention pooling are used in the encoder stage.Graph attention convolution reflects the importance of neighborhood points to the center by calculating the relationship weight of the neighborhood area.Graph attention pooling merges the spatial distribution information of the neighborhood of the sampling points into the features of the sampling points.Experimental results show that the method effectively improves the recognition and segmentation effect.