Video motion compensation for fatigue crack detection in steel structures

Fatigue constitutes a critical limit state affecting the safety of civil infrastructure. Under repetitive loading, structural members are susceptible to fatigue cracking under stresses much lower than the yield strength of the material. In this study, computer vision-based fatigue crack detection using a short video stream taken from a nonstationary camera is presented. Videos taken from a hand-held camera, or an unmanned aerial vehicle (UAV) contain two types of movements: 1) true object movements, and 2) unwanted camera movement due to hand shaking or UAV hovering. In most vision-based structural health monitoring research, feature-based motion compensation techniques are used that require manual selection of fixed objects in the video for feature point selection. Feature point selection from true moving objects in the video could produce inaccuracy in video stabilization. In this study, we propose to use hierarchical model-based motion estimation for global motion compensation, which does not require manual selection of fixed objects. First, we construct a pyramid of target and reference images and then estimate motion from top to bottom of the pyramid while accumulating the geometric transformation, by which the camera movement can be removed. Then, we detect salient feature points in the region of interest and track the motion of feature points throughout the video using the Kanade-Lucas-Tomasi (KLT) feature tracking algorithm. Subsequently, a crack detection, and localization algorithm is applied to search for differential point movements caused by fatigue crack opening and closing. To evaluate effectiveness of the proposed method, a laboratory experiment was conducted on a C(T) specimen with an in-plane fatigue crack. Results show that proposed method was able to effectively compensate the camera motion and detect the presence of the fatigue crack.