Weakly Supervised Point Cloud Scene Segmentation Based On Transformer And Undirected Grap

With the continuous research in mobile robotics,autonomous driving,unmanned aerial vehicles,air and space informatics and other fields,the demand for 3D information is increasing,and 3D point cloud scene segmentation has received more and more attention.In recent years,point cloud scene segmentation based on deep learning networks has made great progress,but most of the methods are limited to the training of noise-free datasets,ignoring the existence of unlabeled or mislabeled category labels in the actual labeling work.Therefore,how to achieve automatic segmentation of noisy point clouds is the key for point cloud data to be used effectively.The difficulties of noisy point cloud scene segmentation technology are mainly reflected in three aspects: first,it is difficult to judge the accuracy of point cloud labels through training in massive data;second,point cloud scene segmentation methods rarely establish multilayer location dependencies;third,it is difficult to solve the problem of pretzel noise in segmentation results.In order to solve the above problems and improve the accuracy of widerange point cloud scene segmentation,this paper proposes a weakly supervised point cloud scene segmentation method based on Transformer and undirected graph,firstly,a weakly supervised point cloud scene segmentation method based on lightweight multi-density Point Transformer is designed to achieve preliminary point cloud scene segmentation,and then a point cloud based on high-order undirected graph network is proposed to Then,we propose a point cloud scene segmentation optimization mechanism based on high-order undirected graph network to solve the pepper noise problem in the preliminary segmentation results.The main work is as follows:(1)For the problem of large volume of point cloud data and difficulty in judging noisy point labels during training,this paper proposes a weakly supervised point noise adaptive learning framework.Firstly,the point cloud data containing noise is input into the deep learning network for training,then point-by-point confidence selection is introduced to discriminate the trusted labels,and finally the cluster-level voting mechanism is used to predict the noise point labels and achieve the label correction of noise points.(2)To address the current problem of insufficient utilization of point cloud context features,this paper designs a lightweight multi-density Point Transformer model.The model obtains multi-layer context dependencies of point clouds through a multi-density sampling algorithm,firstly uses the farthest point sampling to achieve point cloud density stratification,then obtains the neighborhood information of query points under different density layers to form a multidensity feature matrix,and finally uses a self-attentive mechanism to fuse multi-density features.(3)To address the problem of pretzel noise in segmentation results,this paper proposes an ANOVA guided high-order undirected graph network and applies it to point cloud scene segmentation optimization.The network takes supervoxels as the nodes of the undirected graph,firstly generates high-order factors based on local attributes,then calculates the potential functions of the factors by using the principle of variance analysis,and finally uses the belief propagation algorithm to reason about the undirected graph network and obtains the maximum posterior probability state,so as to improve the accuracy of point cloud scene segmentation.