Vision-based modal analysis techniques have gained attention due to their non-contact，full-field measurement capabilities，making them particularly suitable for the dynamic testing of large-scale or thin-walled structures. However，these techniques often require cameras to be fixed to the ground to avoid coupling with the vibrations of the test structure，a requirement that can be too restrictive in real-world applications. This paper proposes a method to compensate for camera motion using homography transformation，followed by the extraction of the test structure’s movement by applying the dense optical flow method to the stabilized video. The procedure involves transforming the video captured by a moving camera using feature matching algorithms，where a homography matrix compensates for six degrees of camera motions. Several "virtual vision sensors" are selected on the edges of the structure，and their vibrations are estimated using optical flow methods. Structural modal parameters are then extracted from the output-only data using stochastic subspace identification algorithms. The proposed procedure was applied to videos recorded using a moving smartphone to conduct an operating modal analysis of a 2 m cantilevered beam. To validate the procedure，the vision-based analysis results were compared with measurements taken with a Scanning Laser Doppler Vibrometer. The results show an average discrepancy of 0.4% and 11.5% for the first five natural frequencies and damping ratios of the beam，respectively. The mode shapes also show strong correlation between the two measurement techniques，as indicated by the diagonal MAC values greater than 98%. Therefore，the proposed procedure effectively cancels out camera motions and achieves accurate estimation of structural modal parameters.