A Visual Localization Benchmark for UAVs in Complex Multi-Terrain Environments

Unmanned Aerial Vehicle self-localization systems that match UAV and satellite images have become widely adopted. These systems demonstrate strong resilience in complex environments and against electromagnetic signals, often serving as an alternative to Global Navigation Satellite Systems (GNSS). A crucial aspect of UAV self-localization systems is the ability to retrieve satellite image using images captured by the UAV. While datasets for this purpose exist, they are primarily focused on lowaltitude urban landscapes with limited terrain variety, hindering their effectiveness in complex environments. Therefore, this paper presents a new dataset, ComplexUAV, specifically designed for the task of UAV self-localization in challenging environments. ComplexUAV adopts high-resolution UAV images that collected from high altitudes with dense sampling, encompassing various terrains such as city, farmland, port, loch, and sands etc. Over 10k UAV and satellite images have been collected and annotated, establishing a new benchmark for UAV self-localization tasks. We also propose a simple but efficient contrastive learning framework with the InfoNCE loss, incorporating dynamic distance sampling strategy and spatial attention module (SAFA) to enhance the model's robustness. Additionally, we introduce a new evaluation metric, MDGM, which is designed to assess the generalizability of models trained on diverse datasets. We extensively evaluate the performance of various feature extraction networks across different training frameworks and datasets. The results show that our framework achieves superior Recall@1 on the University, SUES, DenseUAV, and ComplexUAV, with 95.52%, 99.45%, 94.04% and 88.60% respectively. Furthermore, the model trained on the ComplexUAV achieves the highest generalizability scores on MDGD. The code and dataset will be made available on https://github.com/cjl-2000/ComplexUAV.