Underwater 3D reconstruction via uncertainty-aware binocular-sonar fusion and local transformation refinement

Underwater three-dimensional reconstruction is essential for environmental perception, obstacle avoidance, and intervention tasks performed by underwater unmanned vehicles. Acoustic and optical sensing are the two primary means of underwater environmental perception. However, binocular cameras are sensitive to light attenuation and scattering, whereas imaging sonars suffer from low spatial resolution and missing elevation information. Acoustic-optical fusion can provide more accurate and reliable perception information in short-range underwater scenes. However, existing fusion methods remain limited by ambiguous acoustic-optical correspondences, residual sonar-camera calibration errors, heterogeneous measurement uncertainties, and sparse acoustic observations. To address these problems, this paper proposes an uncertainty-aware three-dimensional reconstruction method that fuses a binocular camera with an imaging sonar. The method operates in a single-pose setting using one synchronized binocular image pair and one sonar image. First, candidate 3D positions are generated for each sonar feature by sampling the unobserved elevation angle and projecting them onto the camera image plane. Then, feature-wise local transformation refinement reduces projection errors by dynamically updating acoustic-optical correspondences, and optical and sonar depths are fused according to their uncertainties. Finally, local surface expansion propagates reliable depth corrections to neighboring geometrically consistent pixels to improve reconstruction completeness. Experiments were conducted in three representative underwater scenes, including a ship propeller blade, an underwater vehicle, and an underwater rock wall. The proposed method achieved coverage rates of 58.09%, 46.37%, and 51.21%, respectively, outperforming the comparison methods in all scenes. In a representative refinement case, the mean acoustic-optical depth residual was reduced from 0.2628 m to 0.1233 m. The results demonstrate that the proposed method improves point-cloud coverage and geometric consistency while maintaining competitive local reconstruction accuracy.