Research On Point Cloud Classification And Segmentation Based On Deep Learning

As a kind of high-precision data,point clouds are widely used in three-dimensional scene fields such as computer vision,autonomous driving,and so on.In recent years,with the rise of deep learning,the application of deep learning to point cloud classification and segmentation tasks has attracted widespread attention from academia and industry.Compared to traditional methods,deep learning based methods can achieve better results by using more data to participate in model training.In point cloud classification and segmentation methods based on deep learning,the method of directly processing point clouds can avoid many problems compared to the method of indirectly processing point clouds,and is a hot research topic.Researchers call it point based method.This article focuses on the point based method,improves the problems existing in the point cloud classification and segmentation model from different perspectives,and propose models with stronger point cloud classification and segmentation capabilities.The main work is as follows:(1)A point cloud classification and segmentation model based on graph convolution and three-dimensional directional convolution is proposed,and graph convolution neural network is applied to the field of point cloud classification and segmentation.The model first treats a point cloud as a node on a graph,calculates its K-nearest neighbor for each node,establishes an edge between two nodes in a local K-nearest neighbor neighborhood,and parameterizes edge features through a graph convolution neural network to capture the local positional relationship between nodes and update the central node feature;Then,a direction coding module is used to divide the neighborhood of a node into eight orientations of fine grained neighborhood blocks,and the node features within the local neighborhood structure are sequentially mapped into different fine grained neighborhood spaces in accordance with the direction of the three-dimensional spatial coordinate axis to extract direction information between nodes,and two direction coding modules are superimposed to increase the receptive field of the network,Improve the robustness of the model for sparse point cloud data and obtain local neighborhood multiscale features.(2)A point cloud classification and segmentation model combining dynamic convolution kernel and dual pooled channel attention is proposed.The existing deep learning point cloud classification and segmentation models usually use shared multilayer perceptrons to learn the characteristics of each point cloud,and reduce the training parameters of the network by sharing the weight sharing mechanism of the shared multilayer perceptrons.However,this mechanism limits the ability of the model to capture different spatial information;In addition,existing models lack the filtering operation for feature channel dimensions after embedding point cloud features into high dimensions,resulting in models that learn the characteristics of all channel dimensions without discrimination and cannot specifically learn the characteristics of channel dimensions that contain important information,which is not conducive to feature learning.To solve the above problems,this model uses adaptive convolution based on dynamic convolutional kernels to replace the shared multi-layer perceptron to solve the problems caused by the weight sharing mechanism.On this basis,a channel attention module combining maximum pooling and average pooling is also designed to enable the model to extract significant features while emphasizing important features while suppressing edge features,thereby more targeted learning of feature information from different channel dimensions.Work(1)Research on the method of graph convolution neural network modeling point clouds,and improve the defect that direction information between point clouds is not considered in the process of graph structure modeling.Use three-dimensional directional convolution to capture direction information between point clouds,enrich local area features of point clouds,and better achieve classification and segmentation.Work(2)Using data driven dynamic convolution to replace the task of shared multi-layer perceptron for feature extraction avoids the problem that shared multi-layer perceptron limits the ability of the model to capture different spatial information.At the same time,a dual pooling channel attention module is designed,which combines the salient features of point clouds obtained by pooling functions and attention mechanisms to enable the model to more targeted learn features of different importance levels and enhance the model’s ability to extract features.Experiments have proved that the above work has achieved good results.