A Comparison of Computer-Vision-Based Structural Dynamics Characterizations

As a specific modern non-contact sensing technology, optical/video information is getting more and more attention employed to interpret structural responses and system status awareness. By means of processing the acquired video, a full-field system information is available which may be applied later to Experimental Modal Analysis (EMA), Structural Health Monitoring (SHM), System Identification (SI), etc., while at the same time, there is no influence to the structural testing such as mass loading and stiffness change. There are numerous technologies to extract the dynamic response of structures from acquired videos. In this paper, several point tracking algorithms are particularly compared, including Lucas-Kanade tracker, Hungarian registration algorithm and particle filter. These computer vision algorithms are implemented to extract the natural frequencies of a lab-scale structure, and the efficiency of each method is investigated regarding the consistency in estimating the natural frequencies and computational time. The recorded video contains external noise caused by lighting change during the experiment, as well as the intrinsic uncertainty on the photosensitive devices. Therefore, the natural frequencies estimated via different algorithms will have different values. An overall comparison between several computer vision algorithms are made in this paper in terms of precision, and computational load.