AI-Powered Human Activity Recognition for Real-Time Detection of Inactive Workers in the Workplace

This research presents a novel way to identify idle employees (non or low productivity) that has not been used in the prior literature utilizing Human Activity Recognition (HAR) in a YOLOv11-CNN (Convolution Neural Networks) detection framework. The model was trained and validated using three datasets from RoboFlow Universe which focused on workplace scenarios and was specifically designed to identify "active" and "sleeping" modes of employee non or low productivity. The main findings show that the model demonstrated a mAP@50 of 97.8% (Dataset 1) with average accuracy of 89.3%and peaked at a F1-score of 0.89, associated with a confidence threshold of 0.569 (Dataset 3). Importantly, the precision and recall curves displayed and no confusion matrices indicated consistent model estimating across confidence thresholds and class-wise performance. Overall, the training displayed smooth convergence and slight overfitting demonstrating potential to transfer into the real-world workplace to monitor employee productivity.