基于环查询和通道注意力的点云分类与分割

点云数据的特征处理是机器人、自动驾驶等领域中三维物体识别技术的关键组成部分，针对点云局部特征信息重复提取、点云物体整体几何结构缺乏识别等问题，提出一种基于环查询和通道注意力的点云分类与分割网络。首先将单层环查询和特征通道注意力机制进行结合，减少局部信息冗余并加强局部特征；然后计算法线变化识别出物体边缘、拐角区域的高响应点，并将其法线特征加入全局特征表示中，加强物体整体几何结构的识别。在ModelNet40和ShapeNet Part数据集上与多种点云网络进行比较，实验结果表明，该网络不仅有较高的点云分类与分割精度，同时在训练时间和内存占用等方面也优于其他方法，此外对于不同输入点云数量具有较强鲁棒性。因此该网络是一种有效、可行的点云分类与分割网络。

Point cloud classification and segmentation based onring query and channel attention

Feature processing of point cloud data is a key component of 3D object recognition technology in robotics, autopilot, and other fields. In order to address the problems of repeated extractions of local feature information of point cloud and lack of recognition of the whole geometric structure of point cloud object, a point cloud classification and segmentation network based on ring query and channel attention was proposed. First the single-layer ring query was combined with the feature channel attention mechanism to reduce local information redundancy and strengthen local features. Then the high response points of the edges and corners of the object were identified by calculating the normal changes, and the normal features were added to the global feature representation, thereby strengthening the recognition of the whole geometric structure of the object. Compared with many point-cloud networks on ModelNet40 and ShapeNet Part datasets, the experimental results show that the network not only has higher accuracy for point cloud classification and segmentation, but also outperforms other methods in training time and memory consumption. In addition, the network is strongly robust for the number of different input point clouds. Therefore, the proposed network is an effective and feasible network for point cloud classification and segmentation.