Objective Autonomous underwater vehicle (AUV) recovery is a critical phase of marine missions, requiring high-precision relative positioning during the terminal approach to the recovery cage. However, large attitude deviations frequently cause visual beacon visibility degradation, leading to severe performance decline or even complete failure of pose estimation. To address this issue, a novel binocular visual guidance method integrating optimized light array layout and pose estimation is proposed.
Method The proposed method consists of two synergistic innovations. First, a "3+3" dual-layer non-coplanar light array layout is designed. On each of two parallel depth planes, three light sources are arranged in an equilateral triangle, and the two triangles are rotated 60° relative to each other. This spatial configuration ensures that at least three non-collinear feature points remain visible even under extreme yaw or pitch angles, fundamentally enhancing the diversity of spatial geometric constraints. Second, the binocular visual pose estimation algorithm is improved. Building upon traditional stereo triangulation combined with singular value decomposition (SVD) for initial value estimation, a nonlinear optimization step based on binocular reprojection error is introduced. The objective function jointly minimizes the reprojection errors of 3D points projected onto both left and right camera image planes, and further incorporates a depth consistency loss term that leverages the known fixed depth differences between corresponding light source pairs in the world coordinate system. The Levenberg-Marquardt algorithm is employed to solve this optimization problem.
Results Comprehensive simulations and ground-based experiments were conducted. Within an operating distance of 0.5 m to 10 m and a yaw angle range of ±60°, the translation error is strictly controlled within 4% of the Euclidean distance between the AUV and the recovery cage, while the rotation error remains below 3.2°. Ablation studies confirm that the depth consistency loss term effectively suppresses point cloud structure drift caused by viewpoint changes and image noise. Even when only three light sources remain visible due to occlusion or extreme viewing angles, the algorithm maintains stable pose estimation. Comparative studies demonstrate that the pose solvability probability of the proposed dual-layer layout is significantly superior to that of single-layer layouts with six or eight co-planar light sources. Simulation comparisons under identical algorithmic conditions further confirm that the dual-layer non-coplanar layout substantially outperforms both single-layer configurations in terms of translation and rotation accuracy.
Conclusion The proposed method significantly improves the robustness and accuracy of AUV autonomous recovery terminal guidance under complex viewing angles through the combination of spatial depth layout and optimization algorithm, providing effective technical support for engineering applications.
DownLoad: