A Deep Learning Approach to Analyzing Active Working Hours of Employees

Abstract

It is conceivable to see how active each individual is at work by observing their various physical postures and gestures. With conventional systems, it is hard to constantly monitor every employee and make the best use of them. Our thesis' major goal is to determine the most effective strategy to utilize each employee's available work time. The challenge of object detection in visual data has been demonstrated to be almost fully resolved by deep learning, a method that mimics the information flow of the human mind. Allowing computers to distinguish not just things but also activities is one of the upcoming key issues in computer vision. In this work, the possibilities of deep learning are examined for the particular challenge of activity recognition in office settings. Data from CCTV footage of several offices during business hours was gathered in order to implement the ‘YOLOv7 object detection’ model. The research implemented a re-labeled dataset of distinct office worker motions to distinguish between employees' levels of activity. With the assistance of Flask, we can build a single-page website, deploy our model, and perform round-the-clock monitoring. After training, the model displays higher accuracy(95.7%), demonstrating how ideal it is for this situation. By employing this approach, it is possible to estimate an employee's productivity by observing their numerous motions while the workplace is in operation. Any official context can use our concept approach, and deployment services can provide ongoing monitoring.