UAV camera pose estimation based on multiscale features and contextual information

Accurate camera pose estimation is essential for unmanned aerial vehicle (UAV) operations, as it underpins reliable navigation, localization, collaborative coordination, and autonomous decision-making in complex environments. However, conventional multi-view geometry methods often encounter difficulties under uneven feature distribution, motion blur, occlusion, and the absence of known camera intrinsics, which significantly limit their robustness and applicability in real-world UAV scenarios. In this work, a UAV-oriented camera pose estimation framework is proposed, integrating a multi-scale FPN-SuperPoint feature extractor, a Local Context-Convolution Glue(LCCGlue), an improved camera motion reconstruction algorithm with simultaneous self-calibration of intrinsics based on the Image of the Absolute Conic, and a pixel-level direction loss optimization module. Extensive experiments demonstrate that the proposed FPN-SuperPoint achieves consistently higher repeatability across illumination, viewpoint, and scale-change scenarios, with particularly strong improvements under scale variations. For homography estimation, the FPN-SuperPoint + LCCGlue combination achieves the highest recall (92.1% @1 px) and correlation (0.895), providing a more balanced and reliable performance than existing approaches. In camera pose estimation, the integration with Pix Adjustment further improves accuracy, with AUC values reaching 71.78 at the 1 cm threshold, while maintaining stable performance even when camera intrinsics are unknown. These results confirm that the proposed framework not only delivers competitive accuracy, robustness, and computational efficiency but also offers practical potential for real-time UAV applications in navigation and mapping tasks.