Point cloud classification and segmentation based on local feature aggregation and driving point cloud sampling

Aiming at the problem that most deep learning models lack the ability to obtain local features, a point cloud classification and segmentation network (FCA-Pointnet ++) based on local feature aggregation and driven point cloud sampling was proposed. A local feature aggregation module was designed and a Feature channel attention (FCA) module was constructed to enhance relevant local features and suppress irrelevant features. A driver point cloud sampling network (DPN) which could change the distribution of initial sampling points according to different tasks was designed. A channel-wise gate module, which could extract local features more accurately and refine features adaptively, was constructed. In the classification task, FCA-Pointnet ++ achieved an overall accuracy (OA) of 95.3% and 93.4% on ModelNet10 and ModelNet40, respectively. In the segmentation task, the mean intersection over Union(mIoU) of FCA-Pointnet ++ reached 86.1%, and the effectiveness of FCA-Pointnet ++ in classification and segmentation tasks was verified by experiments.