Abstract

Sensor-based human activity recognition (HAR) is often deployed under subject shift, uncertain predictions, and tight resource budgets. We introduce CALM- HAR, a confidence-calibrated, lightweight fusion framework designed for reliable on-device inference. CALM-HAR couples a tiny temporal encoder (shallow 1-D CNN) with tree-based learners (Random Forest, XGBoost) and applies post-hoc temperature scaling to obtain well-calibrated class probabilities. A confidence-weighted late-fusion head combines learners in a data-dependent manner, using predictive entropy to modulate each model’s contribution without adding heavy meta-networks. We evaluate under subject-independent (leave-one-subject-out) and cross-dataset (train on UCI HAR, test on PAMAP2) protocols and report accuracy, macro-Precision/Recall/F1 together with Expected Calibration Error (ECE), negative log-likelihood (NLL), and CPU-bound latency/size. Across settings, CALM-HAR maintains competitive macro-F1 while consistently improving probability calibration relative to un-calibrated stacking and single-model baselines, yielding dependable decisions at low cost. The paper details pre-processing, hyper parameters and evaluation settings to support transparent replication. These results indicate that careful calibration and confidence-aware fusion offer a practical alternative to heavier architectures for robust, edge-oriented HAR.

Authors

Kamal Narayan Budholia, Ankur Pandey
LNCT University, India

Keywords

Human Activity Recognition, On-Device Machine Learning, Probability Calibration, Leave One Subject Out (LOSO), Cross-Dataset Generalization