Synthetic Data for Evaluating the Visual Tracking of Construction Workers

Construction sites are hazardous environments due to the sites’ complex and dynamic nature. Hazardous situations especially arise when construction workers collaborate with heavy machinery in the same working space. Pedestrian workers may be missed by the machine operators or get into blind spots. Such hazardous situations can lead to serious incidents accompanied by injuries or even fatalities. Computer vision approaches detecting pedestrian workers in a machine’s surrounding can be used to assist machine operators and prevent accidents. However, collecting data for training and evaluation purposes is difficult. While acquiring a sufficient amount of training samples of pedestrian workers requires high effort, intentionally exposing workers to hazards for the purpose of gathering realistic evaluation data is unwarrantable. In this paper, we apply an approach which detects and tracks pedestrian construction workers in the pivoting range of a tower crane. The detector is trained on a real-world dataset. Since recording realistic test scenarios in the real world emerges to be difficult, we also consider data generated in a 3D environment for evaluation purposes. We compare the performance of our approach on synthetic and real data. Our results show only minor deviations of up to 8% and 3.6% in the spatial accuracy and the track’s length, respectively. This demonstrates that synthetic datasets constitute a reasonable alternative for the evaluation of computer vision approaches on hazardous construction site scenarios.