Research On 3D Point Cloud Semantic Segmentation Based On Attention Mechanism

Three-dimensional point cloud segmentation is the process which divide a point cloud into multiple point sets with the same semantics labels.Because point cloud data itself has high redundancy,uneven sampling,lack of texture,etc,Semantic segmentation of point cloud is extremely challenging.Recently,deep learning has shined in the field of three-dimensional research,which has promoted the evolution of technology and the in-depth research.In the field of three-dimensional semantic segmentation,various deep learning methods have flourished,but the existing methods still fail to make full use of scenes or models.Global dependencies and semantic associations between categories are ubiquitous.For this reason,this paper starts from the attention mechanism to smooth and enhance features by designing a relational learning network layer,thereby improving the saliency of features in each category.This paper proposes two methods which are from the shallower to the deeper,and finally a point cloud semantic segmentation prototype system is constructed based on these two methods.The main results of this paper include the following three points:(1)Progressive 3D scene segmentation method based on self-attention mechanism3D indoor scene point cloud segmentation is to segment the original scene represented by the point cloud into point clouds with semantic tags.In order to alleviate the problem of unbalanced point cloud categories that are common in large-scale indoor scenes,and make full use of the hierarchical nature of the scene itself,we propose a progressive segmentation strategy to divide the scene into ”background” and ”foreground”,we develop a self-attention module suitable for submanifold sparse convolutional networks that is designed to model global dependence and context correlation,which enhances the characterization ability of features.By using the network framework,a corresponding semantic label can be assigned to the point cloud of each input data.In order to verify the effectiveness of the network framework,we conducted a series of supporting experiments.The experimental results show that the network model proposed in this paper could segments 3D indoor scenes well,and performs well in standard data sets.(2)Three-dimensional general semantic segmentation method based on context priorIn order to make full use of the implicit contextual information exists in the scene and model,clarify the category boundary,and alleviate the problem of boundary feature blur caused by continuous convolution operation,This paper constructs a learnable context prior module to model the dependencies within and between classes as prior knowledge.The similar features are smoothed through the context prior matrix,and different types of features are separated as much as possible to improve the discrimination of different types of features,which lead to the reduction of the model search space,and accelerate the convergence.At the same time,in order to build a global dependency,we develop a self-attention module in the network,which is used to aggregate the features that already have contextual semantic relevance,and strengthen the features twice to improve the segmentation accuracy of the model.The experimental results prove that the context-prior module proposed in this paper has strong scalability and wide applicability,In addition,It has achieved good performance on the model and scene semantic segmentation data sets.(3)Design and implementation of Point Seg 3D point cloud segmentation systemIn order to better carry out the work of 3D point cloud segmentation,we have summarized the work of Chapter 2 and Chapter 3,and successfully integrated them into an interactive system that is convenient for researchers and users.The system is embedded The two network models proposed in this paper which could segment the input point cloud and perform visual comparison at the same time.From the perspective of software design and function realization,we have carried out an all-round elaboration on the system,and carried out a demonstration to show the main functions and actual effects of the system.