Objective To address unstable visual features, insufficient trajectory continuity, and localization drift in close-range structural scenes for underwater robots operating in weak-texture, low-illumination, and suspended-particle scattering environments.
Method A tightly coupled multi-sensor underwater simultaneous localization and mapping (SLAM) method based on domain-adaptive visual enhancement is proposed. In the visual frontend, UFEN is fine-tuned through teacher-student distillation using simulated underwater images from HoloOcean, and sign-bit quantization is employed to convert 256-dimensional floating-point descriptors into 32-byte binary descriptors. Visual reprojection, IMU preintegration, and joint DVL-IMU preintegration constraints are incorporated into local factor graph optimization.
Results On the eight sequences of the TANK dataset, the proposed method achieves the lowest root-mean-square absolute trajectory error on seven sequences. Compared with the tightly coupled underwater acoustic–visual–inertial simultaneous localization and mapping (AQUA) algorithm,, the proposed method reduces localization errors by 30.3%–84.2% under degraded conditions, including low illumination, medium-to-high-speed motion, and large-scale loop trajectories. Moreover, the accuracy improvement increases progressively with the severity of image degradation.
Conclusion The experimental results demonstrate that the domain-adaptive fine-tuning of the visual frontend effectively mitigates the degradation of visual constraints caused by underwater environments, significantly improving trajectory continuity and scale stability in complex, large-scale scenes. This provides a reliable, tightly coupled multi-sensor SLAM solution for underwater close-range structural inspection.
DownLoad: