UAV-Based Real-Time Safety Inspection Using Deep-Learning Techniques for Natural and Artificial Hazard Management

Reconstruction and recovery operations after natural hazards pose a risk to the lives of construction workers. Personal protective equipment (PPE), such as helmets, is essential in avoiding fatal and nonfatal injuries at hazardous sites such as landslide monitoring, tunnelling, and other construction sites. Tracking violations in using PPE is difficult to implement systematically due to the high cost and time involved in enforcing safety protocols. This study introduces a new approach, in which high-altitude image and/or video feed from an uncrewed aerial vehicle (UAV) is combined with convolution neural network (CNN)-based object detection models for real-time monitoring of helmet use on construction sites. A high-altitude UAV-based database consisting of 1,822 images covering three construction sites was developed for this study. The data set was preprocessed, annotated, and formatted to train the object detection model. Then different versions of CNN-based models YOLOv5, YOLOv8, and YOLO-NAS were implemented using the data set to identify and localize (1) person, (2) person with helmet, and (3) helmet. The results of the study suggest that an extra-large version of YOLOv5 achieved the highest accuracy among all models [88%–91% mean absolute precision (mAP) overall]. The findings of the study were compared with those of previous models and studies. The implementation of selected model weights to test the video feed from the UAV indicated a reduction in object detection time without compromising the accuracy of prediction. The proposed research empowers both workers and managers to act promptly. The results indicate that the proposed framework is accurate and reliable, demonstrating significant potential to enhance safety inspection within the construction industry.