Research On Point Cloud Classification And Segmentation Methods Based On Attention Mechanisms

Point cloud processing plays an important role in fields such as autonomous driving,virtual reality,and robotics.With the continuous improvement of point cloud acquisition technology,the scale and complexity of point cloud data are also increasing.Therefore,in recent years,how to use deep learning methods to process point cloud has become a research hotspot.To address the challenges posed by the irregularity and disorderliness of point cloud,researchers have designed corresponding networks to solve these problems.However,compared with mature 2D image processing,there is still much room for improvement in the accuracy of current point cloud networks.Point cloud classification and segmentation are fundamental and important tasks in point cloud processing.In this thesis,we will conduct research on point cloud classification and segmentation based on attention mechanisms.For the classification task,this thesis designs a residual-like classification network based on the distance mask attention mechanism,and for the segmentation task,this thesis designs an attention fusion point cloud segmentation network,whose main contents and contributions are as follows:(1)In the classification task,to address the problem of inaccurate feature extraction in local neighborhoods,a distance mask attention module was designed.This module can weight the features of neighborhood points based on the distance relationship between each point and the center point,effectively improving the expression capability of the feature obtained by summarizing the neighborhood.In the distance mask attention module,to mitigate the impact of redundant points in the ball query neighborhood on the attention mechanism,a masking operation was designed to ensure that each neighborhood point only participates in the calculation of weights and feature aggregation once.Furthermore,to address the issue of utilizing intermediate-layer features,this thesis optimized the network architecture.The network generates the global features for each stage after each large module composed of sampling grouping and feature aggregation,and a residuallike connection was designed to fully utilize these global features,effectively improving the model accuracy.This thesis tested the performance of the classification network on the ModelNet40 dataset.Combined with the aforementioned method,the network achieved a classification accuracy of 93.7% on the ModelNet40 dataset,which is an improvement of1.1% compared to the baseline network,PointNet++.This demonstrates the effectiveness of the proposed classification network.(2)In the segmentation task,to better utilize the feature relationships between similar semantic objects,this thesis introduces the vector self-attention module,which can weight features based on their similarity in the feature space.Additionally,this thesis combines the Multi-Scale Grouping method with multiple feature extractors to effectively improve the model’s representation ability.Similar to the classification task,the segmentation network also generates global features in multiple stages.To better utilize these features in the segmentation task,this thesis designs a global feature fusion module,which can provide more sufficient information for each point’s feature,thereby effectively improving the model’s segmentation ability.This thesis evaluated the segmentation performance of the proposed model on the ShapeNet Part and S3DIS datasets.Combined with the aforementioned method,the segmentation model achieved an m IoU of 86.4% on the ShapeNet Part dataset and 66.8% on the S3DIS dataset.Finally,the thesis showcased the segmentation results on these datasets through visualizations.