多层次架构下融合自注意力的三维激光点云语义分割算法

3D Laser Point Cloud with Self-attention under Multi-level Architecture Semantic Segmentation Algorithms

摘要:

为解决三维激光点云局部特征提取不充分且缺少上下文特征信息融合的问题,提出了一种融合自注意力机制与多层次特征提取架构的MAKNet点云特征提取网络。该网络以三维激光点云数据作为输入,通过引入SAA模块提取点云特征,增加中心点特征值与其邻域点特征值之间关注权重值,抑制稀疏点特征低识别度问题,然后采用多尺度特征提取方式,进行多层点特征提取和融合,并进行点云特征跳跃链接,增加被提取点信息涵盖量。实验结果表明,在公开数据集S3DIS上总体准确率相比于PointNet++由80.1%提升至86.9%,在自建输电线路走廊数据集上总体准确率达到96.4%,证明MAKNet网络在语义分割任务上具有良好的鲁棒性以及较强的泛化能力。

Abstract:

To address the insufficient local feature extraction and lack of context feature fusion in three-dimensional laser point clouds, we propose MAKNet, a point cloud feature extraction network that integrates a self-attention mechanism with a multi-level feature extraction architecture. Taking three-dimensional laser point cloud data as input, MAKNet employs an SAA module to extract point cloud features. It enhances sparse point recognition by introducing attention weights between the central point features and neighboring point features. Furthermore, MAKNet utilizes a multi-scale feature extraction approach to extract and fuse multi-layer point features, followed by point cloud feature skip connections to increase the coverage of the extracted point information. Experimental results demonstrated that on the S3DIS dataset, the overall accuracy of MAKNet was 86.9%, which was an improvement compared to that of PointNet++ (80.1%). On a self-built dataset of transmission line corridors, MAKNet achieved an overall accuracy of 96.4%, showcasing its robustness and strong generalization capabilities in semantic segmentation tasks.