A Low Cost Ppe Detection Workflow for Manufacturing Facilities

Abstract

Ensuring personnel safety is essential for manufacturing operations, especially at work sites which increase exposure to various hazards. Thus, automatic monitoring of compliance/non-compliance with the personal protective equipment (PPE) requirements is crucial. Currently, deep learning-based vision systems demonstrate good performance for a variety of applications. However, such systems often require large amounts of annotated data (as many as hundreds of labeled images) for training and a dedicated workstation with GPU (graphical processing units) for training the model. These requirements are often beyond the means of small and medium enterprises (SMEs). To address this challenge, we introduce a workflow for helmet detection that can be easily deployed across several manufacturing sites. We develop and deploy an integrated workflow for helmet detection. We train the YOLOv3 model for helmet detection with data derived from public datasets. This reduces the deployment costs because of the reduced need for obtaining and labeling data from the test environment. The trained model is robust and demonstrates good performance on unseen data acquired from the test environment. We show that including supplementary data (datasets containing mask-wearing individuals) during training improves performance on both public datasets and unseen test data. Finally, our approach also demonstrates good response times (approximately 1 second).

Keywords: Personal Protective Equipment, Artificial Intelligence, Object detection, Deep Learning, Computer Vision