Research On Point Cloud Processing Methods Based On Deep Learning

Point cloud is a flexible geometric representation that has important applications in many fields,including robot vision and navigation,medical image processing,and more.However,processing point cloud data has always been challenging due to its irregular and unordered nature,which makes it difficult to apply convolutional neural networks that are widely used in image processing.The introduction of point-based deep networks,such as Point Net,has enabled direct processing of point clouds,but these methods have weak characterization of the relationships between local 3D points in order to adapt to the irregular and unordered nature of point clouds.To address these issues,this thesis proposes the following:1.Introducing the Top-K operator to replace the widely used max pooling to aggregate local features of point clouds and reduce information loss.Point Net initially proposed using max pooling to aggregate global features,while the DGCNN network used adaptive average pooling to increase global feature representation.However,when multiple salient features are present in the points being aggregated,retaining only one maximum value point results in significant information loss.To address this issue,this thesis proposes using the Top-K operator to replace the max pooling operation,which,under limited storage costs,retains as many salient features as possible to improve the network’s discriminative ability.2.Proposing an enhanced edge convolution layer based on residual attention mechanism,which enhances the network’s expression of important regions and key feature channels.The introduction of residual structure improves the network’s depth,thereby enhancing its ability to abstractly characterize.3.Proposing a point cloud processing network based on the slicing method,which divides point cloud data into slices with contextual relationships,transforming the irregular and unordered point cloud data into regular and ordered data.Due to the unordered nature of point clouds,existing methods often use symmetric functions to aggregate point clouds in the neighborhood,which results in significant loss of local geometric information.To address this issue,this thesis proposes directly cutting the 3D object along a certain direction into continuous areas without changing the original point cloud data,while adding contextual components and learning the original point cloud’s local geometric features and slice context features through two network branches.This thesis compares and conducts ablation experiments,network depth experiments,and parameter setting experiments on point cloud shape classification and shape part segmentation datasets,based on classification success rate and part segmentation m Io U.Additionally,robustness experiments were conducted to study point cloud data loss.The experimental results show that the proposed method significantly improves the shape classification and shape part segmentation performance of the network.