Vision-Based Safety Monitoring for Hazard Detection in Construction Environments

Abstract

Construction sites present complex and rapidly changing working conditions that can increase the risk of workplace accidents. Manual safety inspections are often labor-intensive and may fail to identify hazardous situations in a timely manner. This paper proposes a computer vision framework for automated construction safety monitoring using multi-object detection and spatial relationship reasoning. The proposed architecture combines an object detection network with a scene-level attention module to identify unsafe interactions among workers, construction equipment, and hazardous areas. The illustrative experimental dataset contains 112,000 annotated images collected from 48 construction sites, covering eight categories of safety-related events, including missing protective equipment, restricted-area entry, and unsafe equipment proximity. The proposed framework achieves a mean average precision of 0.914 at an Intersection over Union threshold of 0.5, compared with 0.867 for a conventional object detection baseline. Hazard identification recall reaches 89.2%, while the average processing speed is 31 frames per second on a dedicated GPU. Additional testing indicates that false safety alerts decrease by 23.5% compared with the baseline. These results suggest that intelligent visual monitoring can support more efficient construction safety management.