PPE detection using Deep Learning models optimized to run on Nvidia Jetson platforms
With technological developments in machinery and equipment, businesses must monitor worker compliance with safety rules, such as wearing hardhats and high visibility vests. Existing surveillance cameras at such plants give some compliance, but regular human monitoring is needed to detect whether a worker is wearing personal protective equipment (PPE), such as a helmet or vest. Due to the manual nature of the procedure, it can be hard to detect and predict dangerous events. This research proposes three deep learning (DL) algorithms to verify PPE compliance of workers in real-time from image/video. We try three different approaches and compare the results. First, the DL model detects a person, hardhat, and vest in a joint object regression. In the second approach, DL recognises hardhat and vest-wearing personnel, where the presence of PPE is an attribute of a single object detector. In the third method, the primary detector model detects just a person, then the cropped image is used by the secondary detector model to identify if the person is wearing a hardhat and/or vest as a second stage object detection. All three models are optimised for efficient operation on embedded platforms such as Nvidia Jetson Xavier NX due to the real time requirements of the project. This paper also discusses real-world challenges such as the availability of high-quality data-sets and the effort of labelling image samples. Furthermore, the available data-sets are typically imbalanced and biased, necessitating a great deal of effort and time to resolve these problems. This research proposes a PPE detection framework that evaluates the efficacy of each of the three DL models, considers the limitations of practical applications, and ultimately settles on the optimal approach.