Nighttime UAV Absolute Visual Localization via Dual-Exposure Feature Fusion and Cross-Domain Adaptation

Abstract

When the global navigation satellite system (GNSS) is unavailable, visual localization based on matching uncrewed aerial vehicle (UAV) images with satellite imagery provides a promising alternative. However, under complex nighttime illumination conditions, UAV images often suffer from severe quality degradation and exhibit significant domain discrepancies with satellite imagery, leading to substantial performance degradation in existing absolute visual localization (AVL) methods in terms of both matching accuracy and localization precision. To address these challenges, this article proposes a nighttime UAV–satellite image matching and localization method that integrates dual-exposure feature extraction with cross-domain adaptive learning. Specifically, a dual-exposure feature fusion network is designed to jointly exploit the complementary information from high- and low-exposure UAV images. A cross-attention mechanism and a complex convolution module are further introduced to enhance the network’s ability to capture informative structural regions and key features under nighttime conditions. In addition, a cross-domain adaptation (DA) framework based on generative adversarial learning is constructed to fine-tune the feature extraction network during training, aligning the feature distributions of nighttime UAV images, satellite imagery, and images from the matching dataset. This strategy improves the robustness of cross-domain matching. Experimental results demonstrate that the proposed method significantly improves matching stability and localization accuracy across various nighttime scenarios and outperforms several existing methods, validating its effectiveness for nighttime UAV AVL.