A Multi-Scale Feature Fusion and Extraction Network Based on Point-Wise Attention for Aerial-Ground Point Cloud Place Recognition

Abstract

Place recognition between aerial and ground point clouds is crucial for collaborative robot systems in complex environments but faces challenges due to significant differences in view, texture, and scale between aerial and ground point clouds, as well as limited overlap areas. This paper proposes a multi-scale feature extraction and fusion network with point-wise attention to address these issues, effectively capturing critical geometric information and reducing feature representation differences. Experiments on benchmark datasets demonstrate the method's superior accuracy and robustness, offering a promising solution for tasks like inspection, mapping, and navigation.