Introduction: The construction field is considered one of the most dangerous industries. Accidents and fatalities take place on a daily basis in construction projects. Globally, different levels of government have implemented strict rules and regulations to protect workers on job sites. However, despite the efforts to implement the rules and regulations, accidents occur frequently. Falling from heights is considered the most common cause of death in construction. This study developed a novel system integrating deep learning and drones to monitor workers in real-time when performing at-height activities.
Method: Specifically, a pre-trained deep learning model was used to detect Personal Fall Arrest System components (e.g., safety harness, lifeline, and helmet). The drone was utilized to take images and videos from the construction site, and the data were relayed to the model to detect safety violations. The system was tested and validated in real construction sites and in a controlled lab environment to verify the model's effectiveness under different light and weather conditions.
Results: The overall accuracy of the system was 90%. The model's precision and recall were 97.2 % and 90.2%, respectively. The average time taken to detect a violation was around 12 seconds.
Conclusions: Moreover, the Area Under Curve - Receiver Operating Characteristics chart showed that the trained model was very good and precise in detecting and differentiating the desired objects.
Practical applications: This fast, reliable, and economical system can aid in saving many lives if implemented and utilized properly in real construction sites.
Keywords: Fall from heights; Machine learning; Personal Fall Arrest System (PFAS); Real-time detection; Unmanned Aerial Vehicles (UAV).
Copyright © 2022 National Safety Council and Elsevier Ltd. All rights reserved.