A Geometric–Temporal Coupling Method for Vision-Based Pose Monitoring of Bridge Erecting Machines Under Degraded Visual Observations

Reliable and temporally consistent pose monitoring of bridge erecting machines (BEMs) is essential for geometric control and construction safety. However, vision-based pose estimation is susceptible to abrupt jumps under long-distance imaging, partial occlusion, and time-varying observation quality. To address these challenges, a geometric–temporal coupling method combining geometric consistency refinement with occlusion-aware temporal state estimation is developed for temporally consistent vision-based pose monitoring of BEMs. Frame-dependent observation confidence derived from visibility, image quality, spatial distribution, and reprojection consistency jointly regulates geometric refinement and temporal correction. The method is evaluated through controlled resolution-degradation simulations using RANSAC-EPnP (RE), robust DLT (RDLT), and Kalman-filtered RANSAC-EPnP (RE-KF) as comparative baselines, together with separate occlusion-degradation simulations, and is further examined through a field application at the Shuimotan 88 + 168 + 88 m continuous rigid-frame bridge. Under the 7000 × 7000-pixel condition, the proposed method achieves the lowest translational errors and the lowest maximum absolute rotational errors, while its mean rotational errors remain comparable to those of RDLT and RE-KF. At 3700 × 3700 and 400 × 400 pixels, the proposed method yields the lowest RMSE, MAE, and maximum absolute differences across all six pose components. The three-axis translational trajectories remain free from sequence interruption under 25%, 50%, and 75% occlusion. In the field test, relative to the Leica TS60 reference measurements, the RMSEs of the vision–total-station differences for the x, y, and z components are 0.7331, 1.2151, and 0.7735 mm, respectively, while those for α, β, and γ are 0.0254°, 0.0113°, and 0.0081°. These results indicate that the proposed method improves pose-estimation stability under the investigated resolution- and occlusion-degradation conditions and demonstrates the feasibility of six-degree-of-freedom BEM pose monitoring in the investigated field scenario.