A Dynamic Graph-based Time Series Analysis Framework for On-site Occupational Hazards Identification

Different factors combined invariably cause construction fatalities at any time, most of which could be avoided if workers followed the on-site regulatory rules. However, compliance of regulatory rules is not strictly enforced among workers due to all kinds of reasons, even after prior education and training. To address the difficulties of on-site safety management, this paper proposes a graph-based time series analysis framework to dynamically integrate visual and linguistic information for on-site occupational hazard identification. A vision-based scene information understanding approach is introduced to process on-site images via a combination of deep learning-based object detection and individual detection, together with a novel dynamic graph structure to represent time-series information for integrated reasoning of hazards identification. As a case study, the hazards of grinder operation were successfully identified in the experiments with high accuracy.