一种面向低采样率的点云数据处理网络

在直接处理点云的三维神经网络中，采样阶段实现了对原始点云中关键点的筛选，对于整个网络的性能及网络的抗噪能力具有重要作用。目前主流的最远点采样（FPS）方法在处理大规模3D点云数据时计算量大且耗时，并且低采样率时经过FPS采样后模型性能下降明显。针对这两个问题，提出一种面向低采样率的点云数据处理网络AS-Net。设计一个新的采样模块代替原backbone中的FPS，其由两个Layer组成，每个Layer基于长短期记忆网络获取原始点云与采样点云之间的联系权重，从而高效提取关键信息，去除冗余信息。在此基础上，利用注意力机制选择特征值较高的原始点云作为采样点，采样点作为后序任务的关键点输入到网络，进一步提高网络模型性能。基于ModelNet40数据集的实验结果表明，在低采样率条件下，AS-Net仍可达到81.6%的分类准确率，与使用FPS作为采样方法的网络模型相比提高52.7%。此外，其对噪声干扰具有很强的鲁棒性，对于大场景的分割时间效率优于同类采样方法。

A Point Cloud Data Processing Network Oriented to Low Sampling Rate

Point-cloud data sampling, grouping and fusion have been applied in several state-of-the-art networks to improve their segmentation and classification results.Among these, the sampling stage maintains the key points of the original point cloud, which plays an important role in the performance and anti-noise ability of the network.However, two defects limit the application of models based on Farthest Point Sampling(FPS) to large-scale 3D point-cloud data:(1) FPS requires considerable computation and time;(2) at low sampling rates, the performance of the model under FPS significantly decreases.To solve those two problems, a point cloud data processing network oriented to low sampling rate, AS-Net, is proposed, which uses a new sampling module to replace FPS in other backbones.The sampling module is composed of two layers:(1) a Long Short-Term Memory(LSTM) network structure;(2) an attention mechanism.LSTM is used as the feature extraction method to obtain the relational weight relative to the subsequent tasks, efficiently extract key information, and remove redundant information.The attention mechanism is used to select the original point cloud with a high eigenvalue point as the sampling point;sampling points are input to the network as key points of the subsequent tasks to improve the performance of the network model.Experiments show that the proposed method still achieves a classification accuracy of 81.6% on the ModelNet40 dataset at a low sampling rate, which is 52.7% higher than that of the same network using FPS.In addition, the proposed method has a strong robustness to noise interference, and its segmentation time efficiency for large scenes is better than that of similar algorithms.