Research On Point Cloud Semantic Segmentation Algorithm Based On Deep Learning

In recent years,with the development of three-dimensional data acquisition devices such as Li DAR,obtaining three-dimensional point cloud data has become more convenient.Threedimensional point cloud semantic segmentation,as a fundamental technology for scene understanding,has become a research hotspot in fields such as augmented reality,intelligent factories,and autonomous driving.Although three-dimensional point cloud data can effectively preserve spatial raw information,it also has characteristics such as geometric invariance,permutation invariance,and non-structural properties.Neural networks are prone to loss of information between points and spatial information when extracting point cloud features,making it difficult to mine deep-level feature information from three-dimensional point cloud data.This paper focuses on the research of point cloud semantic segmentation tasks and the above issues.The specific work is as follows:(1)Due to the correlation between points in three-dimensional point cloud data,it is difficult to extract spatial features from them,and neural networks may suffer from gradient disappearance problems during training.To address these challenges,a point cloud semantic segmentation network based on residual multi-perceptron and attention mechanism was designed.This neural network directly takes three-dimensional point cloud data as input and extracts point cloud features through a feature extraction module.In the feature extraction module of the point cloud,the residual multi-perceptron can pass shallow features to the deep network,which can avoid the gradient disappearance problem that may occur during the training process of the point cloud semantic segmentation network to some extent,thereby improving its learning efficiency.Introducing the attention mechanism in the feature extraction module can effectively focus on the importance between points and enhance the learning ability of the point cloud semantic segmentation network.The neural network was trained and validated on the Shape Net dataset,and the experimental results showed that the segmentation accuracy could reach 85.7%.Compared with the benchmark networks Point Net and Point Net++,this neural network has a faster convergence speed,and the segmentation accuracy is improved by 2% and 0.6%,respectively,showing a good segmentation effect.(2)To address the problem of feature fusion between local point clouds and global point clouds,a point cloud semantic segmentation network based on feature fusion was designed.The neural network extracts global features and local features separately from the point cloud,and then inputs the global feature matrix and local feature matrix into a feature fusion module to output a more expressive point cloud feature matrix,which is used to achieve point cloud semantic segmentation by classifying the point cloud.Considering the interaction between points in local point clouds,the graph convolution method is used for feature extraction in the local feature extraction module.Additionally,a channel spatial double attention mechanism is introduced in this network to give greater weight to key points and mine the spatial geometric features of three-dimensional point clouds.Experimental results show that the overall segmentation accuracy on the S3 DIS dataset and Semantic3 D dataset can reach 85.1% and 91.8%.At the same time,the network reduces the parameter count as much as possible,and accelerates the convergence speed of the network.