【摘要】 点云是三维空间中物体坐标点的集合，是一种能完整表达场景信息的重要数据格式，广泛应用于机器人、计算机视觉、自动驾驶等领域，而点云的有效分类是许多应用的前提。基于PointNet和DGCNN(Dynamic Graph CNN for learning on point clouds)，提出一个用于处理点云分类的PPN(Prototypical Point Network)神经网络模型，在PPN的特征提取模块中，通过点云中点之间的距离关系生成更加精准的局部特征。通过对局部特征进行加权聚合，得到更有代表性的全局特征。在ModelNet40和ShapeNet Parts数据集上用PPN进行点云分类和分割实验，结果表明PPN在点云分类和分割任务中均取得较好的表现。更多还原

【Abstract】 Point cloud is a collection of object coordinate points in three-dimensional space,and it is an important data format that can completely express scene information. It is widely used in robot,computer vision,automatic driving and other fields. Effective classification of point cloud is the premise of many applications.Based on PointNet and DGCNN(Dynamic Graph CNN for learning on point clouds),this paper proposes a PPN(Prototypical Point Network)network for point cloud classification. The key of PPN network is the PPN feature extraction module,which learns the weight between local points through the different distance relationships between points,generates local features,and then continuously aggregates features to extract the features of each point cloud category,better establishes the topological relationship between discrete point clouds,and extracts more accurate local features and global features Experiments on point cloud classification and segmentation using PPN on ModelNet40 and ShapeNet Parts datasets show that PPN achieves good performance in point cloud classification and segmentation tasks.更多还原