Complex background segmentation for noncontact cable vibration frequency estimation using semantic segmentation and complexity pursuit algorithm

This paper proposes a new complex background segmentation method based on the modified fully convolutional network semantic segmentation for noncontact cable vibration frequency estimation. The estimation of frequency from video data is challenged by the presence of background object motion, which directly impacts the accuracy of the video-based method. To address this issue, image tests were carried out among the existing model (U2-Net) to explore the effect of the efficient channel attention (ECA) and convolutional block attention module (CBAM) on cable segmentation performance. As a result, a relative optimal model was identified. This modified model was then used to remove the complex background, while retaining the vibration signals specific to the cable. Subsequently, phase matrices encoding cable vibration were calculated using a phase-based motion estimation algorithm at various cable locations. The modal response of the cable vibration was estimated using the complexity pursuit (CP) algorithm from the segmented video. Finally, the vibration frequency of the cable was estimated. The proposed method was validated on a small-scale cable model. The results are in good agreement with the values sampled by the accelerometer, with an average relative error of 4.50%. This estimation shows the significant potential of the proposed method in structural health monitoring.