Inertial Sensor-Based Hierarchical Contrastive Multi-View Learning for Human Activity Recognition
Francisco M. Calatrava-Nicolás, Vítor Fortes Rey, Paul Lukowicz, Óscar Martínez Mozos · 2025
This paper addresses Human Activity Recognition (HAR) using wearable inertial sensors, where model generalization to unseen individuals remains challenging owing to multiple sources of data heterogeneity. To mitigate this, we introduce a sensor-based hierarchical multi-view framework that defines each view as the set of sensors positioned symmetrically on the body. This framework uses a shared encoder across all views to learn view-invariant representations, while view-specific heads are used to capture view-dependent nuances. We reinforce cross-view consistency and view sensitivity through supervised contrastive learning at two hierarchy levels: the shared encoder and the view-specific heads. Our method yields macro F1 gains of 0.7-12.4 percentage points (pp) over state-of-the-art baselines in multi-sensor classification across four of the five public HAR benchmarks, and ranks second on the remaining benchmark. For single-sensor classification, it improves macro F1 by 1.3-8.5 pp on average over all the evaluated sensors. Further experiments confirm robustness to both sensor displacement and data scarcity. Github code available . https://github.com/FranciscoCalatrava/MultiView-Supervised-Contrastive-HAR.git.