Target-free recognition of cable vibration in complex backgrounds based on computer vision

The cable system is a critical load-bearing member and a vital factor of cable-stayed bridges; its health condition indicates the operation safety of these structures. The cable force estimation based on vibration is a major indicator for assessing the safety of cable-bearing bridges. Contact sensors are used to collect vibration in traditional cable systems. Recently, computer vision-based non-contact remote monitoring has gained popularity due to its cost, efficiency, and safety ad-vantages. However, the cable is a unique structure with poor vision imaging, a low pixel share, and susceptibility to ambient interference. To address these issues, this study proposes a deep learning-based cable vibration recognition system. In complex environments, this system provides reliable recognition of cable vibration without the necessity of markers. The system consists of a composite model based on Resnet-34 (Residual Neural Network of 34 layers) and Swin-B (the base model in Swin Transformer), a linear rigid body motion recognizer based on Hough linear detection, and data processing. The pre-trained composite model performs cable segmentation on the video captured by the camera; this new video contains only cable data and is then transmitted to the linear rigid body recognizer, which recognizes cable vibration in the new video, and the cable force is determined by processing vibration data. As a result of the model experiments conducted in the laboratory and the cable test of cable-stayed bridges under complex outdoor conditions, the effectiveness and robustness of the system have been verified. The error of the proposed method in this study is found to be less than 2.0% comparing the results with that calculated from conventional sensors employed in the tests. The test results indicate that the system proposed in this paper is capable of conducting cable detection in complex environments.