Visual Planar Range Measurement and Uncertainty Evaluation for Unmanned Surface Vehicles Based on Dynamic Homography and Semantic Boundary Localization

To address vision-based ranging for unmanned surface vehicles (USVs) when stable depth sensors are unavailable and to provide declarable measurement uncertainty, this paper proposes a visual planar range measurement and uncertainty evaluation method based on dynamic homography and semantic boundary localization. The method targets planar distances from the USV to water–land and water–obstacle boundaries in the bird’s-eye view (BEV) and constructs a vision-based ranging instrument using a monocular camera and pose sensors. Inertial navigation data are used to estimate USV attitude in real time and dynamically update the homography between the camera and target plane, thereby establishing a measurement model from image pixels to the BEV plane. Lightweight semantic segmentation extracts key boundaries and projects them onto the BEV for range observation. A physically interpretable pixel-domain error model is introduced to characterize anisotropic uncertainty along the two pixel axes and its dependence on range and viewing angle, and an analytical Jacobian-based error propagation scheme is derived to propagate pixel-level uncertainty to the standard uncertainty of planar range. Real-world on-water experiments in straight and corner scenes, with equal-sample acquisition and joint statistical analysis, show improved BEV geometric consistency and reduced ranging error after attitude compensation. Uncertainty consistency is evaluated using normalized residuals, coverage, and calibration metrics, showing structurally consistent uncertainty prediction with mild underestimation that can be corrected by simple scaling. A data-driven sensitivity analysis and quantile-based applicability-boundary description are further provided for engineering parameter selection, and comparisons with representative learning-based monocular depth methods are included as reference baselines.