关键词: 跨模态, 航空照片, 地面点云, 地点识别, 旋转平移不变

Abstract: To address the localization problem of unmanned vehicles exploring unknown areas under satellite-denied conditions, a cross-modal place recognition approach from aerial images to ground point clouds has been proposed. An Air-Ground Collaborative Place Recognition Network architecture, AG-PRNet, was designed for this purpose. Cross-modal place recognition faces significant challenges such as substantial heterogeneity in low-dimensional features across different modalities and noticeable rotation and translation variations.This paper first performs data preprocessing by projecting point clouds into the Bird's Eye View (BEV) space to reduce heterogeneity differences with aerial images. Then, a rotation and translation invariant feature encoding module(RATI-CNN) is designed to extract rotation and translation invariant features from multi-modal data. Finally, a cross-attention module is used to fuse and learn shared features across multi-modal data, effectively enhancing the robustness of cross-modal feature matching.A cross-modal place recognition dataset, CDPR Dataset, was constructed, and comparative experiments and ablation studies were conducted on this dataset using mean recall rate as the evaluation metric. Experimental results show that the proposed cross-modal place recognition method achieved a Top-1 recall rate of 60.08% and a Top-5 recall rate exceeding 76%, demonstrating a clear advantage over other methods.

Key words: Cross-modal, Aerial Images, Ground Point Clouds, Place Recognition, Rotation and Translation Invariance