| description abstract | The tracking of construction resources (e.g., workforce and equipment) in videos, i.e., two-dimensional (2D) visual tracking, has gained significant interest in construction industries. Many studies have relied on 2D visual tracking methods to support the surveillance of construction productivity, safety, and project progress. However, few efforts have been aimed at evaluating the accuracy and robustness of these tracking methods in construction scenarios. The main objective of this research is to fill that knowledge gap. Compared with previous work, a total of 15 2D visual tracking methods were selected here because of their excellent performances identified in the computer vision field. Then, the methods were tested with 2 videos captured from multiple construction jobsites during both day and night. The videos contain construction resources, including but not limited to excavators, backhoes, and compactors. Also, they are characterized by attributes such as occlusions, scale variation, and background clutter. The tracking results were evaluated with the sequence overlap score, center error ratio, and tracking length ratio, respectively. The results indicated that (1) the methods built on local sparse representation were more effective, and (2) the generative tracking strategy typically outperformed the discriminative one when being adopted to track the equipment and workforce in construction scenarios. | |
