Machine Learning for Sensor-Based Human Activity Recognition

Human Activity Recognition (HAR) has become a key enabler in wearable health technology and fitness analytics. This paper presents a machine learning framework for classifying six physical activities—Walking, Running, Cycling, Swimming, Resting, and Yoga—using sensor-derived physiological and motion features. A dataset of 8,000 observations with 14 attributes including heart rate, steps per minute, step entropy, and distance traveled is utilized. A Random Forest classifier (n_estimators=50) is trained on an 80/20 train-test split and evaluated against a Decision Tree baseline. The Random Forest achieves a test accuracy of 85.06% with a macro-averaged F1-score of 0.83, compared to the Decision Tree's 76.5% accuracy. Feature importance analysis identifies heart rate and steps per minute as the most discriminative predictors. Correlation analysis reveals strong relationships (r ≈ 0.95) between distance and calories burned. A real-time prediction interface is implemented to demonstrate practical deployment. Results demonstrate the effectiveness of ensemble learning combined with interpretable feature analysis for robust activity recognition in resource-constrained wearable systems.