Study On Semantic Segmentation Of 3D Point Clouds Based On Point MLP

Three-dimensional point cloud semantic segmentation has important applications in industrial production,medical services and other fields.However,due to the disordered and unstructured characteristics of point cloud data,the requirements for feature extraction of point cloud data are not completely consistent among point cloud segmentation tasks of different scales.Therefore,based on the Point Multi-Layer Perceptron(Point MLP)processing method,the thesis proposes a semantic segmentation model for point cloud components at the fine-grained level,as well as a semantic segmentation model for largefield point clouds with large data processing scale.The main work of this thesis is as follows:(1)Data collection and pre-processing.A robot platform equipped with Li DAR(Light Detection And Ranging)and depth camera was constructed to collect point cloud data.The collected dynamic process data and static file data were converted into a unified point cloud data form.Pre-processing operations such as downsampling and downsampling are performed on the open data sets used in the experiments to facilitate subsequent experimental comparisons.(2)Semantic segmentation of point cloud parts.In order to maintain the fine-grained structural characteristics of point cloud objects,the adaptive farthest point sampling algorithm is proposed for the extraction of key points for each class and part.Secondly,An improved SENet module is proposed,and three different compression operators are designed for comparative experiments to further explore their channel geometric connections.Finally,each point was classified and labeled through feature interpolation to achieve the purpose of semantic segmentation of the overall point cloud object and its parts.(3)Semantic segmentation of large point cloud scenes.The point clouds within complex scenes are widely distributed and have large density differences,which leads to a decrease in segmentation accuracy.To solve the problem,a multi-fusion local region construction method is proposed.This method enabled the model to retain the key information of sparse point clouds while reducing the computing power consumption of dense point cloud areas as much as possible.Meanwhile,a local feature extraction algorithm based on spatial self-attention mechanism is proposed,which not only links the geometric relationships between key points and their local neighbors,but also extracts the geometric relationships between local geometric points.(4)Experimental results evaluation.The thesis uses intersection ratio and overall accuracy metrics to evaluate the segmentation results.Experimental results show that semantic segmentation model for point cloud parts can achieve an 85.4% m Io U(mean Intersection over Union)for parts and 82.8% m Io U for classes on the dataset ShapeNetCore,and can maintain stability and accuracy when the channel dimension is halved.The semantic segmentation model of point cloud large scenes can achieve an overall accuracy of 87.1%on the dataset S3 DIS,with a m Io U of 66.7%,effectively improving the segmentation accuracy.