Structured Correspondence Fusion for Image-Driven UAV Navigation
Abstract
In the image-driven UAV navigation task using the PairUAV dataset, we aim to estimate relative heading and signed translation range between a source-target image pair for last-meter pose alignment. This task faces challenges including drastic viewpoint/scale changes, repeated scenes, low-texture areas and periodic heading ambiguity. We present a visual-geometric fusion framework integrating global appearance features and structured correspondence geometry. A shared ResNet50 extracts common embeddings for two images, with absolute feature differences capturing pairwise appearance gaps. SuperGlue matches are encoded into an 11-channel spatial displacement map and a 40-dimension global descriptor to jointly model local and global motion. The fused representation predicts heading via cosine-sine parametrization and range via bin classification. Given the lack of an official validation partition in PairUAV, we adopt a deterministic group-based split to mitigate training-test overlap. Our system achieves a score of 0.0853 on the official evaluation server (official baseline: 0.6340; lower scores mean superior performance). The source code is publicly available at https://github.com/zhuiguangzhe-zz/UAVM2026-GISP-HEU.