Camera-Based Human Activity Recognition for Productivity Monitoring using LSTM

This paper proposes a camera-based Human Activity Recognition (HAR) framework for real-time productivity monitoring in workplace environments. The system uses MediaPipe Pose to obtain skeletal landmarks, a Convolutional Neural Network (CNN) for spatial feature encoding, and a Long Short-Term Memory (LSTM) network for temporal modeling to recognize activities such as typing, reading, writing, phone use, and idle behavior. A synthetically generated dataset of office actions was used for training, yeilding 98% training accuracy and 96% validation accuracy, and the TensorFlow Lite version runs in real time at 20–25 FPS. The lightweight model supports efficient on-device deployment while maintaining strong performance. A weighted productivity mapping module translates recognized activities into an engagement index from 0 to 1. Experimental results show that the CNN–LSTM–TFLite pipeline provides accurate, fast, and privacy-preserving activity recognition suitable for non-intrusive workplace productivity assessment.