Hierarchical activity classifiers and live CNN inference using chest-worn accelerometer recordings.
Designed and compared hierarchical classifiers for clinical activity monitoring.
Used 12 Hz 3-axis MEMS accelerometer data from a chest-worn sensor.
Extracted handcrafted time- and frequency-domain features.
Used Python, Random Forest, XGBoost and SVM.
Classification structure: Stage-1 coarse: 4 groups; Stage-2 fine: 12+ activities.
Deployed a pre-trained CNN for live inference on multi-hour recordings.