Viewpoint-Aware Pose Estimation Framework for Cooperative UAVs

Pose estimation from monocular vision is essential for UAV applications, yetexisting methods often struggle in real-world settings. Traditional techniquesbased on markers or hand-crafted features are computationally efficient but unre-liable in cluttered or unstructured environments. Learning-based approaches,while powerful, typically demand extensive target-specific annotation andretraining, limiting their generalizability. This paper proposes a robust, key-point training-free pose estimation framework that leverages pre-trained visualcorrespondence models—SuperPoint and LightGlue—to eliminate the need fortask-specific keypoint detection or pose regression. Target localization is per-formed using an off-the-shelf YOLO detector, followed by viewpoint-awaretemplate matching to discretize target appearance under varying views. Withinthe detected region, SuperPoint features are matched via LightGlue, and anovel Coverage Score evaluates the spatial distribution of correspondences toreject degenerate configurations prior to PnP–RANSAC pose recovery. AnUnscented Kalman Filter integrates asynchronous measurements for tempo-rally stable yet responsive estimates. The framework requires no target-specificretraining, enabling seamless deployment across diverse targets and environ-ments. Extensive evaluations—including indoor, outdoor, and visually degradedscenarios—demonstrate robust, consistent performance and significant reduc-tions in annotation effort and system integration complexity, outperformingstate-of-the-art learning-based methods in various conditions.