Abstract
Accurate image correspondences are essential for geometric inference in vision-based UAV sensing because camera pose, 3D reconstruction, and spatial measurement depend on the geometric validity of matched points rather than on match count alone. This study develops a hierarchical Fourier–Mellin/phase-correlation framework for sequential RGB imagery acquired by a UAV camera sensor. Global similarity registration is followed by recursive local refinement with spectral, forward–backward, overlap, and radiometric acceptance criteria. Performance was assessed on 338 held-out stereo pairs using independently bundle-adjusted camera geometry as an external reference and a common relative-orientation pipeline for all methods. The proposed method achieved 63.91% matching success and 60.06% reference-valid success, compared with 89.05% and 27.81%, respectively, for Area-LSM. For successfully oriented pairs, median relative-rotation and baseline-direction errors were 0.0322° and 3.7987°, respectively. On common successfully oriented pairs, the proposed approach yielded lower errors than Area-LSM and ORB, whereas SIFT, Harris, Förstner, Moravec, and FAST generally retained higher success rates and lower errors. Baseline-direction accuracy was strongly dependent on parallax and homography dominance. These results support independent sensor-geometry validation and geometry-aware hybrid matching for UAV vision-based sensing.

This publication has 23 references indexed in Scilit: