Crowded Scene PPE Detection Using Attention Based YOLOv7 and Alpha Pose

Personal Protective Equipment (PPE) plays a significant role in ensuring the safety of workers during operations. Safety regulations in each company mandate the use of various pieces of equipment, such as gloves, helmets, vests, and safety shoes, to minimize the risk of injuries. Deep learning techniques are employed to monitor and detect whether workers are wearing PPE correctly. This paper proposes enhancing PPE detection in crowded scenes by leveraging attention-based YOLOv7 and AlphaPose for human pose estimation. Our approach uses human pose estimation to verify whether the detected equipment is positioned correctly. This step enables the removal of false detections, contributing to an increase in detection accuracy. We address the limitations observed in existing works, particularly their need for more accuracy in crowded scenes. Our approach involves AlphaPose, which demonstrates higher accuracy in human pose estimation and is better suited for crowded scene detection. We compared our results with YOLOv7 pose estimation and ViTpose, using mean average precision as the evaluation metric. The result shows that AlphaPose achieved the highest mean average precision (mAP), 0.883 and 0.601, at the IOU threshold of 0.5 and 0.95, respectively.