基于几何仿射与注意力的三维激光点云分类

三维点云分类和分割对于三维重建和自动驾驶等技术的发展具有积极的推动作用。三维点云数据具有无序、不规则和稀疏等特点，因此三维点云分类和分割的研究面临诸多挑战。PCT分类网络采用标量注意力机制提取三维点云局部特征，具有良好的三维点云特征学习能力，在三维点云分类和分割任务中表现出先进的分类精度。然而PCT在对三维点云数据进行下采样时忽视了其稀疏性对几何结构所产生的影响，从而无法充分地提取局部特征致使三维点云分类和分割精度下降。针对该问题，本文提出一种基于注意力机制的三维点云分类分割网络GAM-PCT，具体地，GAM-PCT网络采用了向量注意力机制对单通道特征的权重进行调节，利用减法关系和邻域位置编码对三维点云邻域求取注意力特征，同时在对整体点云下采样时插入即插即用的几何形状仿射（GAM）模块来解决三维点云局部区域的稀疏性问题，进而提升网络的分类准确率。实验结果表明，与PCT三维点云分类和分割网络相比，所提出GAM-PCT网络在数据集ModelNet40上的分类精度提升了0.3%，而在ScanObjectNN数据集上的分类精度提升了1.9%，在ShapeNet数据集上的分割平均交并比值提升了0.2%。同时在网络参数量和FLOPs指标上分别降低了0.31 G和0.69 M。实验结果表明改进后网络的复杂度得到了简化，充分验证了改进方法的有效性。

Point cloud classification based on geometric affine and attention mechanism

The classification and segmentation of 3D laser point cloud have a positive role in promoting the development of 3D reconstruction and automatic driving technology. The 3D laser point cloud data has the characteristics of disorder, irregularity, and sparsity, so the research of 3D laser point cloud classification and segmentation faces many challenges. The point cloud transformer (PCT) classification network uses the scalar attention mechanism to extract the local features of 3D laser point clouds. It has a good 3D laser point cloud feature learning ability and shows advanced classification accuracy in 3D laser point cloud classification and segmentation tasks. However, when PCT downsamples the 3D laser point cloud data, it ignores the influence of its sparsity on the geometric structure, so it cannot fully extract the local features, resulting in the degradation of the classification and segmentation accuracy of the 3D laser point cloud. To solve this problem, this paper proposes a three-dimensional laser point cloud classification and segmentation network GAM-PCT based on the attention mechanism. Specifically, the GAM-PCT network uses the vector attention mechanism to adjust the weight of the single channel features and uses the subtraction relationship and neighborhood location coding to obtain the attention features of the three-dimensional laser point cloud neighborhood, At the same time, a plug and play geometric affine (GAM) module is inserted to solve the sparsity problem of the local area of the three-dimensional laser point cloud when downsampling the whole point cloud, thereby improving the classification accuracy of the network. The experimental results show that, compared with the PCT three-dimensional laser point cloud classification and segmentation network, the classification accuracy of the proposed GAM-PCT network on the data set modelnet40 is increased by 0.3%, while the classification accuracy on the ScanObjectNN data set is increased by 1.9%, and the average intersection ratio of segmentation on the shipment data set is increased by 0.2%. At the same time, the network parameters and the flops index are reduced by 0.31 g and 0.69 m respectively. The experimental results show that the complexity of the improved network is simplified, which fully verifies the effectiveness of the improved method.