Segmenting Personal Protective Equipment Using Mask R-CNN

Every workplace must emphasize safety. The health of an organization’s workforce directly affects its effectiveness. Workplace safety is the culmination of regulations, strategies, and protective measures intended to reduce risks, accidents, and other forms of harm in the workplace. Making sure all employees are wearing the right safety gear, such as hard hats and high-visibility clothing, is one way to mitigate these risks. In this research, we employ computer vision and deep learning to detect, localise, and segment hardhats, safety jackets, and employees from an image. The images used were first pre-processed with computer vision to achieve consistency and eliminate any unwanted elements that could hinder output efficiency. In order to segment our classes in an image, the deep learning section uses Mask R-CNN and compares the effectiveness of a ResNet-101 and ResNet-50 backbone. The pretrained MS COCO dataset was expanded to include two new classes– hardhats and safety jackets–and the model was assessed using the Jaccard Index and mAP. The ResNet-101 Mask-RCNN model outperformed the ResNet-50 model with mean JAC 88.60% and average mAP 73.01%.