In-Air Experimental Validation of Geometry-Prior-Guided Lamp-Center Localization for AUV Docking

Localization of circular guidance lamps under saturation and oblique viewing is critical for AUV docking, yet existing methods underuse the known beacon geometry and suffer when detected centers are displaced. This study presents a geometry-prior-guided localization method validated through entirely in-air experiments. Candidate luminous contours are selected using single-lamp and array-level consistency derived from the known four-lamp layout. Contour points are refined to subpixel precision, and residual projective bias is corrected by inferring direction and sign from projected geometry and estimating magnitude from ellipse deformation. On 417 rectified images across 21 pose groups, the method achieves a mean localization error of 5.732 pixels, 25.54% below the average of seven comparison methods, and yields the lowest image-level error on 377 images. In planar pose estimation, the resulting centers produce mean translation and reprojection errors of 16.074 mm and 5.794 pixels, reductions of 12.01% and 24.94% respectively, with a mean rotation error of 2.662°, revealing a trade-off relative to the strongest baseline. These results indicate that geometry-guided localization can reduce center-estimation errors across the tested in-air imaging conditions. The current CPU implementation processes approximately 3.62 frames/s; therefore, real-time onboard deployment remains a subject for future work.