This paper presents a visual localization system for underwater robots, aimed at achieving high-precision relative positioning and dynamic target tracking. A 3D AprilTag reference structure is constructed using a cubic configuration, and a high-resolution camera module is integrated into the AUV for real-time tag detection and pose decoding. By combining multi-face marker geometry with a fused state estimation strategy, the proposed method improves pose continuity and robustness during multi-tag transitions. To address pose estimation discontinuities caused by viewpoint changes and tag switching, we introduce a fusion-based observation-switching Kalman filter, which performs weighted integration of multiple tag observations based on relative distance, viewing angle, and detection confidence, ensuring smooth pose updates during tag transitions. The experimental results demonstrate that the system maintains stable pose estimation and trajectory continuity even under rapid viewpoint changes and frequent tag switches. These results validate the feasibility and applicability of the proposed method for underwater relative localization and tracking tasks.

Conventional cooperative control methods for multi-AUV systems typically rely on quasi-steady hydrodynamic assumptions and do not explicitly account for time-varying uncertainties in ocean dynamics. In addition, controller parameters are often tuned empirically. As a result, under complex disturbed flow fields and communication constraints, AUV swarms are prone to group fragmentation and reduced polarization, which undermines stable cooperative navigation. To address these limitations, we propose a double-gyre-flow-optimized autonomous underwater vehicle swarm (DGF-OAS) model for coordinated operations in time-varying flow fields. The proposed model incorporates a heading-aware graph attention mechanism to adaptively adjust adjacency weights among agents with different roles. It further integrates the Lennard–Jones potential to preserve safe inter-vehicle spacing and embeds a periodically varying double-gyre flow field to characterize ocean disturbances. Bayesian optimization is then employed to automatically identify suitable weights for the alignment and attraction–repulsion terms, thereby improving swarm cohesion and environmental adaptability. Simulation results demonstrate that, under flow-field disturbances, DGF-OAS achieves group polarization of up to 96%, reduces the average task completion time by 15.84% compared with the baseline model, and attains a task completion rate of 97%, significantly outperforming the compared methods. These findings indicate that the proposed approach exhibits strong adaptability and stability in complex environments and offers an effective solution for AUV swarm control.