A Human Activity Recognition algorithm using three wearable IMU-based units

We present a wearable human activity recognition system based on a three-unit inertial measurement network, with sensors on the thorax, abdomen, and lower back. Data were collected from ten healthy participants performing five activities—sitting, standing, walking, running, and cycling—in a controlled indoor protocol. Feature vectors were built from statistical and frequency-domain descriptors and pruned via Pearson correlation (>0.95) to ensure independence. Two classifiers were trained with a subject-wise split (8 training+ 2 test subjects, 45 combinations) and internally optimized via LOSO cross-validation. K-Nearest Neighbors achieved 75.1 % accuracy, 0.748 precision, 0.751 recall, 0.732 F1-score, and a macro-AUC of 0.936. The Random Forest classifier outperformed KNN with 84.2 % accuracy, 0.841 precision, 0.853 recall, 0.838 F1-score, and macro-AUC of 0.970. Confusion matrices revealed the greatest misclassification between sitting and standing—two static, vertical postures with similar IMU signatures. These results demonstrate that a low-power, flexible BSN with three IMUs can achieve robust classification performance, paving the way for comprehensive, real-world activity monitoring.