Target-free Full-Field Modal Identification of Bridges Based on Computer Vision

Abstract

Full-field mode shape is critical for bridge condition assessment, as it can capture minute distortions induced by local degradation. However, existing computer vision-based measurement methods often fail to balance implementation efficiency and spatial resolution due to their reliance on artificial targets, typically measuring responses only at sparse and discrete points. This paper proposes a targetfree full-field modal identification framework for bridges based on Recurrent All-Pairs Field Transforms (RAFT). Firstly, dense optical flow is estimated from naturally distributed surface textures in video sequences. Secondly, a displacement extraction procedure is used to obtain physically consistent displacement responses for modal identification. To balance field of view (FOV) and local spatial resolution, a segmental stitching strategy with overlapping FOVs is then introduced for global mode shape reconstruction. Finally, laboratory experiments on a simply supported beam are conducted for validation. The identified displacement responses agree well with the measurements obtained from eddy current displacement sensor. Compared with KLT and Canny-Zernike, it yields smoother and more accurate continuous mode shape. For the first mode, the MAC and RMSE relative to the reference result are 0.998 and 0.042, respectively. The segmental stitching strategy therefore provides an effective way to reconcile FOV and resolution in modal identification.