Machine learning analyses for dynamic images recorded from physical model tests

This study aims to measure vibrational displacement of structures in physical model tests using high-speed recorded images and an open-source machine learning–based model, YOLOv5. Two dynamic experiments were performed, and their images were recorded: a vibrated block on a shaker and a single-degree-of-freedom structure on dynamic centrifuge tests. For the shaker experiment images, four parameters were examined: different training methods, pre-trained models, partial area patterns, and pattern types. The patterns include black-and-white checks, circles, squares, crosses, and X shapes. For the different training methods, training with one labelled image and its copies showed better performance than training with all images. The pre-trained model analysed other videos small errors although the errors increased with greater camera-to-structure distances. Among the types of patterns, the X pattern performed the best, with ∼2% errors and an average coefficient of determination of 0.9928. The images from the dynamic centrifuge test were analysed by the machine learning code and compared with results from two other popular object-tracking software programs. The tracking results from the machine learning model showed performance comparable to that of the other image-based tracking programs. The results suggest that YOLO-based image tracking can be effectively applied to laboratory vibration tests.