Hierarchical Adaptive Normalization: A Placement-Conditioned Cascade for Robust Wearable Activity Recognition

Keywords: human activity recognition, domain adaptation, sensor orientation correction, on-device learning, adaptive batch normalization

TL;DR: We propose a lightweight, placement-conditioned adaptive normalization cascade that improves wearable activity recognition under sensor variability, achieving robust real-time performance with minimal overhead.

Abstract: Wearable Human Activity Recognition (HAR) systems suffer from performance degradation due to sensor placement and orientation variability. We propose a hierarchical adaptive cascade that first normalizes sensor orientation using a gravity-based correction and infers coarse placement context via signal variance analysis. A novel stability gate prevents adaptation during unstable dynamics, while a subsequent placement-conditioned adaptive Batch Normalization refines feature representations. Evaluations on public and custom dynamic-activity datasets demonstrate a consistent improvement in macro F1-score over static models and complex unsupervised domain adaptation approaches, all while maintaining low latency and minimal memory overhead. These results expose real-world pitfalls in conventional approaches and highlight the promise of our adaptive method for on-device HAR.

Submission Number: 110