Improving 2D displacement accuracy in bridge vibration measurement with color space fusion and super resolution

Abstract

Accurately measuring vibration displacement in bridge structures is crucial for structural health monitoring. Traditional sensor-based methods often limited by high costs and restricted measurement ranges, while vision-based methods provide a simpler, cost-effective alternative for long-distance, non-destructive vibration measurements. Nevertheless, previous visual methods have typically relied on a single color space and overlooked feature preservation during upsampling, resulting in inaccurate detection of small-amplitude targets. This paper proposes a bridge vibration displacement measurement method that integrates multi-color space information with a super-resolution reconstruction module. First, a Color Space Fusion Module converts RGB images to YCbCr and fuses information from both spaces, achieving a balance between performance and parameter count with an increase of less than 0.1M. Next, traditional upsampling is replaced by a learnable super-resolution module to preserve feature information. Finally, a penultimate module integrates global information to improve measurement accuracy before outputting detection results The measurement accuracy and the fit of regressed vibration displacement curves with standard data were validated for the proposed method on beam-like structures in laboratory settings as well as on real large-span bridge structures, with various image acquisition angles and focal lengths. In laboratory settings, the proposed method showed a 6% improvement in mAP over the baseline model YOLOv9c, with a reduction of approximately 54% in average MAE in the time domain across three detection targets, and a 55% reduction in average RMSE. Additionally, on real bridge structure, the proposed method achieved an average MAE of 2.97 and RMSE of 3.84 Compared to existing visual measurement methods, the proposed approach demonstrated superior performance under these diverse conditions.