SCC-Loc: A Unified Semantic Cascade Consensus Framework for UAV Thermal Geo-Localization

A new computer-vision framework called SCC-Loc can pinpoint a drone’s position using only a thermal camera, achieving a mean localization error of 9.37 meters without GPS, according to research published on arXiv [1]. The system, detailed by Xiaoran Zhang and collaborators, addresses a long-standing problem for Unmanned Aerial Vehicles (UAVs) operating in Global Navigation Satellite System (GNSS)-denied environments, where traditional satellite-based positioning is unavailable [1][2]. Cross-modal Thermal Geo-localization uses a drone’s thermal imagery matched against a pre-existing satellite map, but the profound visual differences between thermal and visible-light images have historically introduced severe feature ambiguity that corrupts standard matching pipelines [2]. SCC-Loc, which stands for Semantic-Cascade-Consensus localization, unifies the entire process under a single DINOv2 backbone shared between global image retrieval and fine-grained matching via MINIMA$_{\text{RoMa}}$ [1][2]. The design reduces the system’s memory footprint while enabling zero-shot absolute position estimation [2]. The framework introduces three modules to overcome the thermal-visible modality gap. A Semantic-Guided Viewport Alignment module adaptively optimizes satellite crop regions to correct initial spatial deviations [1][2]. A Cascaded Spatial-Adaptive Texture-Structure Filtering mechanism enforces geometric consistency to eliminate dense cross-modal outliers [1][2]. Finally, a Consensus-Driven Reliability-Aware Position Selection strategy derives an optimal location through physically constrained pose optimization [1][2]. To train and evaluate the system, the researchers constructed Thermal-UAV, a dataset containing 11,890 diverse thermal queries referenced against a large-scale satellite ortho-photo and a spatially aligned Digital Surface Model [1][2]. The dataset was built to address a scarcity of public benchmarks for this specific task [2]. In experiments, SCC-Loc established a new state-of-the-art, delivering a 7.6-fold accuracy improvement within a strict 5-meter threshold over the strongest existing baseline [1][2]. The work, along with its code and dataset, has been made publicly available on GitHub [2].