Target-Free Stereo-Vision-Based Modal Identification of a Cantilever Beam Using Physics-Aware Deep-Feature Tracking

Reliable target-free vision-based modal identification remains challenging because accurate image correspondence does not necessarily guarantee physically reliable structural measurements. This study presents a Physics-Aware Adaptive Deep-Feature Tracking (PA-ADFT) framework that integrates learned feature extraction, stereo-temporal correspondences, physics-aware measurement validation, and triangulation to reconstruct three-dimensional camera-coordinate displacement trajectories for structural dynamic analysis. The framework was experimentally validated under controlled laboratory conditions through five repeated free-vibration tests on a single healthy aluminum cantilever beam at 10 mm tip-release amplitude, using synchronized stereo cameras and strain gauges. PA-ADFT maintained 7–9 accepted structural measurement points with 90–95% median tracking coverage across five experiments and accurately identified the first bending natural frequency with a median value of 7.02 Hz, compared to 7.03 Hz from strain measurements and 7.06 Hz from the finite-element model. A staged comparison showed that this frequency is already recovered by SuperPoint–LightGlue–Lucas–Kanade tracking; the subsequent geometric, temporal, and kinematic gates preserve track coverage and three-dimensional point identity. Calibrated stereo triangulation recovered the camera-coordinate trajectory in three dimensions, preserving the dominant first-mode bending behavior in the vertical displacement component. Vertical displacement was validated quantitatively against reference measurements; the depth component was reconstructed but not independently validated. Together with physics-aware measurement checks, this supports target-free stereo vision as a viable method for identifying first-mode bending in a single-specimen, single-amplitude laboratory setup.