Research on 3D Point Cloud Segmentation based on KCA- PointNet Network

Currently, mainstream 3D point cloud segmentation methods still have shortcomings, which have an impact on the accuracy of point cloud segmentation. In particular, by using the MLP layer with shared weights, PointNet, and its improved version, generate redundant channel features when mapping 3D point cloud data to high-dimensional space. This study intended to improve the PointNet network and introduce the key point attention and channel attention mechanisms into the PointNet model to build the KCA-PointNet network. Taking the point cloud data generated via oblique photography as the experimental data, the experimental results show that 1) after adding the KCA module, the segmentation accuracy of buildings, parks, and other places is greatly improved; 2) the segmentation accuracy of the combined use of the KA and the CA module is about 3% higher than that when they are used independently; and 3) the different connection modes of the KA and the CA modules also have an impact on the accuracy, and upon connecting the two modules in series, the data first pass through KA and the segmentation effect after CA is better than that of the other two connection methods. Through the improvement of the PointNet network, this study not only eliminated the redundant channel features, but also extracted the key points in the point cloud, reduced the possibility of losing important information in the process of point cloud data sampling, and effectively improved the accuracy and efficiency of point cloud segmentation. Furthermore, the results of this study will help to broaden the application scope of deep learning in the field of point cloud data, promote the development and innovation of point cloud segmentation algorithms, and lay a scientific foundation for deep learning point cloud segmentation.