A robust phase-based vibration perception pipeline using general curvelet transform and adaptive data fusion strategy

Vision-based measurement techniques have attracted widespread interest across engineering fields. Phase-based motion estimation has been widely used due to its sub-pixel precision and high-resolution sensing capabilities. However, noise introduced during different calculation processes, such as image noise, gradient calculation noise, and unstable phase interference, can compromise the accuracy of local phase extraction and subsequent vibration displacement estimation. Addressing these noise-related challenges with a single approach remains difficult. In this study, a structural vibration perception pipeline based on phase-based motion estimation is designed to improve the accuracy and robustness of motion estimation under various types of noise environments. Specifically, multiscale local phases are extracted using the general curvelet transform, which provides near-optimal sparse representations. Multiscale local amplitudes are integrated into an adaptive data fusion strategy that eliminates phase-unstable scales through self-evaluation indices, thereby avoiding reliance on fixed thresholds. Meanwhile, the spatial frequency map, used as another input for data fusion, is estimated via a double filtering approach followed by the O’Shea refinement algorithm to reduce noise in the estimated values. To demonstrate proof-of-principle, a numerically simulated two-dimensional complex-valued image and two-motion simulated videos with varying image noise were performed. It demonstrated that the proposed method outperformed several existing algorithms in terms of vibration estimation accuracy. Validation experiments were conducted on both rigid and flexible structures, with comparisons of estimated displacements confirming the superior accuracy and robustness.