A Forewarning Method for Falling Hazard from Hole Based on Instance Segmentation and Regional Invasion Detection

Falls from height (FFH) are a substantial type of accident in the construction industry and cause immense fatal injuries and asset losses. Skills training and hazard awareness training cannot effectively solve the root cause of the fall-related hazard. Implementing automation and informatization with technical intervention to prevent construction hazards has been recognized as a focus topic. However, the recent literature appears to lack scientific management research concerning prevention technology of fall accidents caused by workers reach the hazard zone of the hole. This paper proposes a computer vision-based approach to contribute to the topic of fall hazard prediction and forewarning related to holes in construction sites: (1) instance segmentation module based on Yolact for hole detection and virtual fence generation (2) object detection module based on YOLOv5 for worker detection, and (3) regional invasion detection module for behavior detection. The results show an accuracy rate and a recall rate of behavior detection with value of 69% and 74%, respectively. By detecting the occurrence of workers’ hazard behaviors approaching the hole in real-time, the causal chain of accidents is controlled through forewarning to prevent accidents happen and provide effective assistance for the construction safety administrator.