Deep Learning-Based Safety Assurance of Construction Workers: Real-Time Safety Kit Detection

With high advancements in technology, the death rate in the construction sector is still the highest compared with different sectors. The safety kit/personal protective equipment (PPE) has also been improved for the safety assurance of workers on site, however, workers also forget to wear the proper PPE several times. Manual inspection for the workers’ safety is not feasible for a large site, moreover, it would be inefficient as numerous site workers could not be checked simultaneously on a labor-intensive basis. It is furthermore a prime task to provide a safer environment to the laborer on site. This motivates the authors to propose an automated approach using computer vision and deep learning for real-time PPE detection, i.e., helmets, vests, and masks (HVM) monitoring. The authors also created a dataset via manual data collection of images and videos including six different classes which are helmet, vest, mask, safe, partial safe, and not safe class. The safe class would ensure the complete safety of having workers wearing all three major PPE, and the not safe class on the other hand would illustrate the absence of these PPE. The dataset includes 5,000 images and their respective annotations for the six classes. The dataset includes 5,000 images and their respective annotations for the six classes. With the modified architecture of the you only look once (YOLO), an object detection technique has been implemented for the detection of the PPE (HVM). The proposed YOLO-V6 technique outperforms in comparison with the previously used SOTA YOLO-V5 which is used by several researchers. YOLO-V6 yields better accuracy in a shorter interval of time and attains the maximum mean average precision (mAP) of 88% using NVIDIA Tesla T4 and (52 FPS) speed in contrast with the predecessor algorithm.