Contrastive Learning Based Visual Place Recognition Pre-Training Framework for UAV Geo-Localization

With the rapid advancement of remote sensing technology, the Visual Place Recognition (VPR) of Unmanned Aerial Vehicles (UAV) based on satellite images has drawn interest in the space-air-ground research area. However, traditional VPR methods have encountered significant challenges caused by the inherent image data heterogeneity of sensors on UAVs and satellites, weather conditions at the time of image capturing, and seasonal changes on ground targets. For better VPR accuracy and robustness, this paper proposed a pre-training framework based on the contrastive learning method, leveraging large-scale unlabeled remote sensing data. Furthermore, we have constructed a pairs aerial-to-satellite image dataset to fine-tune and evaluate the performance of the downstream VPR method. Experimental results demonstrate that this pre-training framework can effectively improve the VPR performance from R@1 accuracy from 0.9479 (Resnet 50 backbone without pre-train) to 0.9701 (Resnet 50 backbone with pre-train).