太赫兹科学与电子信息学报2024，Vol.22Issue(7)：730-740,11.DOI:10.11805/TKYDA2022101
基于注意力机制的三维点云目标识别
Attention mechanism based 3D point cloud target recognition
王阳 1肖顺平1
作者信息
- 1. 国防科技大学 电子科学学院,湖南 长沙 410073
- 折叠
摘要
Abstract
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.关键词
三维点云/目标识别/注意力机制/深度神经网络Key words
3D point cloud/target classification/attention mechanism/deep neural network分类
信息技术与安全科学引用本文复制引用
王阳,肖顺平..基于注意力机制的三维点云目标识别[J].太赫兹科学与电子信息学报,2024,22(7):730-740,11.
