Research On The 3D Point Cloud Segmentation Method Based On Deep Learning

The segmentation of three-dimensional point cloud data is a key step in scene understanding in the field of computer vision which is the hot spot of Scientific research scholar.With the continuous progress of point cloud acquisition technology,the application of threedimensional point cloud data has gradually penetrated into various fields such as automobile driving,smart home and military defense,the application prospect is very broad.After briefly analyzing the current situation of point cloud segmentation,the theoretical basis of point cloud segmentation is introduced in the paper,including the basic concepts of point cloud,the definition of point cloud segmentation and the point cloud segmentation network based on deep learning.This paper designs a joint semantic-instance segmentation network based on deep learning for complex indoor scenes,while completing the semantic segmentation and instance segmentation of 3D point cloud data.The multitask learning backbone network controls the network volume,minimizing the network parameters,based on the realization of multitask processing.The joint semantic feature and instance feature module breaks the segmentation accuracy bottleneck of the existing semantic feature space and the instance cloud segmentation network,obtain higher segmentation accuracy and achieve a win-win situation for the two tasks.Aiming at the poor ability of the existing networks to extract the local feature information and difficulty in capturing the point-level context relations of the point cloud data,the multitask learning network is added approximate 3D convolutional operations for learning the rich local feature information of the point cloud in this paper.The network constructs a nonlinear function as a kernel function that enables convolution operations on point cloud data,enhancing the extraction of local feature information from point-gathered midpoints.The structure sensing loss function is adopted for the instance segmentation problem of 3D data,while sensing the geometric information of the object space and the instance embedding information,so that the network has a better supervision and discrimination ability of instance segmentation.For the problem of not realizing high-precision segmentation in the point cloud segmentation network in complex indoor scenarios,the feature fusion module and the semantic feature joint instance feature module are designed in this paper.The feature fusion module connects multiple layers by jumping to fuse the features of different levels respectively,so as to strengthen the integration of the network into the information contained in the data.Joint Semantic feature and instance feature network increases the semantic and instance discriminative features through multitask level features union,improving the accuracy of semantic segmentation and instance segmentation of point clouds.Finally,complex comparative validation experiments are carried out on the proposed point cloud a joint semantic-instance segmentation network based on deep learning.Large indoor scene dataset S3DIS and fine component segmentation dataset ShapeNet are selected for comparison and ablation experiments to verify the excellent performance of the proposed network.