Reliable structural vibration measurement from image sequences requires accurate recovery of small vibration induced motion while maintaining spatial consistency across the vibrating component. Many deep optical flow methods rely on locally sampled correlation cues and contextual features during iterative refinement, with limited explicit guidance on global structural consistency. Such a limitation may result in fragmented flow fields and spurious local motions when vibration amplitudes are small. A Semantic Conditioned Context Feature Fusion (SCoFF) module is developed to introduce global structural information into the flow estimation process. Structural priors extracted by a semantic encoder are transformed into channel wise affine modulation parameters and an adaptive gating coefficient, enabling selective recalibration of contextual features through a gated residual pathway. The resulting semantic guidance improves motion consistency within structural components while preserving local matching information for accurate displacement estimation. Experimental results demonstrate improved measurement accuracy over representative optical flow methods. At a displacement amplitude of 0.4 mm, an RMSE of 0.055 mm and an NRMSE of 6.561% are achieved, representing the lowest measurement errors among the evaluated methods and confirming reliable performance for structural vibration measurement.

Deep optical flow, semantic features, stereo vision, vibration measurement