Describing Construction Hazard Images Identified from Site Safety Surveillance Video

Construction accidents are a major cause of occupational fatalities globally. On-site hazard identification is crucial to prevent such accidents. CCTV is commonly used for safety surveillance on construction sites, and can be utilized for machine learning-based automatic hazard identification. A Construction Hazard Description System (CHDS) was developed in this study to systematically label site objects and describe hazard scenarios. CHDS builds on the ontology of Taiwan Occupational Safety and Health Administration (TOSHA) for hazard classification and construction accident risk scenarios. The system produces site images and associated hazard descriptions that can be used to train automated construction accident risk identification systems through machine learning. According to domain experts, CHDS is effective in assisting construction safety personnel in describing hazard images collected on site, achieving high accuracy rates in both attribute description and hazard classification. It is concluded the system has great potential in improving the task of captioning construction hazard images.