【摘要】 针对现有三维点云语义分割方法从点云中提取的建筑物存在漏提取与目标不完整的问题，本文提出一种融合注意力机制的多尺度特征聚合点云建筑物提取方法。其中，设计双重注意力机制的局部特征提取模块加深中心点和邻域点关联，利用全文感知聚合模块从广泛的角度捕获全局信息，并通过低阶语义特征和高阶语义特征的深入融合来提高三维点云特征细化的有效性和高效性。基于公开数据集与人工标注的密集匹配点云数据集的实验结果表明，本文方法能够有效提高建筑物提取精度。更多还原

【Abstract】 This paper proposes a multi-scale feature aggregation point cloud building extraction method that integrates attention mechanism to address the problems of missing buildings and incomplete targets extracted from point clouds by existing 3D point cloud semantic segmentation methods. Among them, the local feature extraction module with dual attention mechanism is designed to deepen the correlation between the center points and neighboring points. The global context aggregation module is employed to comprehensively capture global information from a broad perspective, and the deep integration of low-level and high-level semantic features enhances the effectiveness and efficiency of refining 3D point cloud features. The experimental results based on publicly available datasets and manually annotated dense matching point cloud datasets show that the proposed method can effectively improve the accuracy of building extraction.更多还原