一种基于空间特征注意力机制的点云分析方法

Point Cloud Analysis Method Based on Spatial Feature Attention Mechanism

针对现有的基于点的网络平等地对待所有的点从而无法有效关注重要特征的问题,在激光雷达点云处理领域引入注意力机制,即CSA模块,其中CA表示通道注意力,SA表示空间注意力。两个模块以数据驱动的方式自动学习不同特征通道信息和不同空间位置信息的重要性,从而提升网络在点云分类和分割任务上的表现。在基于点的网络中引入了上述两个模块,提出了CSA-PointNet++结构。实验结果表明:所提方法在ModelNet40数据集上的分类准确率达93.20%,在ShapeNetPart数据集上的部件分割实验的平均交并比(mIoU)为82.62%,优于其他对比方法,验证了所提网络的有效性;同时,在真实世界自建数据集上,所提方法的分类准确率达92.14%,证明了网络在真实世界的数据上具有良好的泛化能力。

To address the limitations of existing point-based networks,which treat all points with equal emphasis,thereby overlooking crucial features,this paper introduces an attention mechanism to lidar point cloud processing.This mechanism,referred to as the CSA module,integrates the channel attention and spatial attention elements.In a datadriven approach,the two proposed modules autonomously learn the importance of different feature channel information and spatial location information,thereby enhancing the performance of the network on point cloud classification and segmentation tasks.This paper introduces the two modules stated above in a point-based network and proposes a CSAPointNet++architecture.The results reveal that the proposed method achieves an accuracy of 93.20%for classification experiments on the ModelNet40 dataset and a mean intersection over union(mIoU)of 82.62%for part segmentation experiments on the ShapeNetPart dataset.This performance is better than that of other comparative methods,indicating the effectiveness of the proposed network.Moreover,classification experiments of the proposed method on a real-world self-constructed dataset yield an accuracy of 92.14%,demonstrating the excellent generalization capability of the proposed network on real-world data.