Point Cloud Sampling And Sample Balancing Based On Local Geometric Constraints

As an important research direction in point cloud research,point cloud classification has made significant progress in recent years due to the continuous efforts of researchers,demonstrating outstanding performance in some scenarios.Point cloud processing methods represented by PointNet extract features through multi-layer perceptrons with shared weights,and use the method of multi-layer perceptrons and max pooling symmetric functions to infinitely approximate any geometric structure,achieving competitive results in terms of performance.However,point cloud data is easily affected by external factors during the acquisition process,and there is also a problem of imbalanced class samples.Traditional point cloud sampling methods are susceptible to the quality and distribution of point clouds,and may lose some important local detail information.In this paper,we focus on the research of deep learning-based point cloud sampling methods,and the main work is as follows:1.In response to the problems of traditional sampling methods such as FPS and Learning To Sample,such as being highly sensitive to the quality and distribution of the input point cloud,losing local detail information,and failing to capture local geometric structural features,this paper proposes a point cloud sampling neural network based on local geometric constraints.This network extracts more important and representative points by extracting the features of local geometric structures,thereby retaining points with richer information in the original point cloud.To verify the effectiveness of the sampling results,this paper compares the point cloud recognition results of different combinations of S-NET,FPS,and LGPR with PointNet,PointNet++,and DGCNN.Experimental results show that the proposed method can more effectively extract local geometric features from the original point cloud,resulting in more representative point cloud samples.This method successfully improves the classification accuracy of the simplified point cloud,proving its effectiveness in classification tasks.2.The proposed method aims to address the issues of imbalanced sample sizes in the dataset and the incompatibility of deep learning-based point cloud sampling with downstream tasks.A generative network-based sample balancing technique is proposed to supplement the imbalanced samples in the dataset.Then an end-to-end training strategy is applied to the downsampling network using an upsampling method on the downsampled results,resulting in a sampling network that is compatible with downstream tasks.Experimental results show that the downsampled network obtained through end-to-end training can effectively enhance the representativeness of the original point cloud,and can improve the quality of input data for downstream tasks more effectively than the FPS method.Furthermore,the augmented dataset can effectively improve the classification accuracy of the test dataset under unchanged network models and sampling methods.This demonstrates that the proposed method has good robustness and applicability and can produce good results in practical applications.3.Based on the theoretical foundation mentioned above,the requirements and feasibility of systematic functions were analyzed,and the principle of simplicity and ease of use,as well as software design specifications,were used as guidelines to design and implement point cloud preprocessing,classification,and training and testing models based on the PyQt5 framework.This was done to verify the feasibility and effectiveness of the proposed method in this paper.