Hot Work Control Measures Using CNN-Based Object Detection and Projective Geometry for Industrial Surveillance Application

Industrial safety management has been a common challenge for many industries to implement since industrial hazards could cause fatal risks and unscheduled downtime. In this paper, we proposed an alternative approach for hot work control measures using CNN-based object detection and projective geometry, which could be integrated with the existing surveillance system. This method aims to monitor hot work activity and implement the risk assessment policy, which could control by the hazard area control. The dataset for our study consisted of 909 images of hot work activities captured by two closed-circuit television (CCTV) cameras. There are two steps to our methodology, which are the object detection stage and the bird's-eye perspective transform stage. In the first stage, Workers, Welders, and Hot works are localized using an object detection algorithm, which is YOLOv5. To maximize the F1-score performance of object detection, we ran the experiments to train YOLOv5 with three levels of augmentations: low, medium, and high. For the second stage, four points are required in the method of transforming the object's Cartesian coordinates into the new coordination in a bird's-eye perspective. The radius distance threshold has to be manually calibrated for each specific camera point of view. If there is a worker that moves into the hot work radius, the violation alarm is triggered. The results show that medium augmentations produce the best results, with an overall mAP and F1-score of 0.77 and 0.74, respectively. In addition, the predefined distance threshold is also required and can vary in the different scenarios in the bird's-eye perspective transform stage.