基于注意力机制的三维点云目标识别

Attention mechanism based 3D point cloud target recognition

针对现有基于深度学习方法的三维点云目标识别算法存在多层感知法缺少点间的特征交互、对点云间欧式距离的依赖、未考虑特征通道层面关联性问题,提出一种基于注意力机制的三维点云(PAttenCls)目标识别算法。采用基于点的空间注意力机制,挖掘各点之间的注意力值,实现自适应的云邻域选择;同时采用基于点的通道注意力机制,给特征通道自适应分配权重,实现特征增强。此外,在网络中添加了一个几何均匀化模块,以应对不同局部区域几何结构的不同特征模式。所提算法在ModelNet40数据集上的识别准确率为93.2%,在ScanObjectNN数据集最难子集上的识别准确率为80.9%,并在实测数据上验证了算法的有效性。实验证明了本文所提算法可以更好地提取点云的特征信息,使点云识别结果更加精准。

In response to the issues with existing 3D point cloud object recognition algorithms based on deep learning methods,such as the lack of feature interaction between points in multi-layer perceptrons,reliance on Euclidean distance between point clouds,and failure to consider the correlation at the feature channel level,we propose an attention mechanism-based 3D point cloud(PAttenCls)object recognition algorithm.The spatial attention mechanism based on points is employed to explore the attention values between points,achieving adaptive neighborhood selection for point clouds;meanwhile,the channel attention mechanism based on points adaptively assigns weights to feature channels,enabling feature enhancement.Additionally,a geometric uniformization module is added to the network to address the different feature patterns of different local regions'geometric structures.The proposed algorithm achieves a recognition accuracy of 93.2%on the ModelNet40 dataset and an accuracy of 80.9%on the most difficult subset of the ScanObjectNN dataset,and its effectiveness is verified on real-world data.Experiments have proven that the proposed algorithm can better extract feature information from point clouds,making the point cloud recognition results more accurate.