Conceptual Design of Human Detection via Deep Learning for Industrial Safety Enforcement in Manufacturing Site

Abstract

Industrial workers are vulnerable to hazard and accidents. There could be many factors that contribute for these to occur including human error. Standard operating procedure and safety guideline have been set up to be followed by the workers with manual supervision where total adherence is required in a wide range of operation and hence, often lead to inefficiency. Thus, this work has proposed a preliminary work on safety monitoring within the potentially danger area to make the process to be efficient and reduce the manual supervision burden via deep learning. This work has adopted YOLO network for feature extraction and human detection in several monitoring areas. Then, counting module is executed to retrieve the data of how frequent the monitoring area is being interrupted. Prior to that, a region of interest (ROI) would be set up where human is detected only in the ROI. Lastly, measure the area of intersection between human and ROI to decide whether the subject is in the monitoring area or vice versa. The number of counts indicates the risk of accidents occur in the monitoring area. The higher the counts, the higher the risk in that region. This conceptual design can be extensively used in many ways for safety monitoring as it requires less supervision and becomes a safety measure by enforcing industrial safety in manufacturing sites.