# Agile human activity recognition for wearable devices based on online incremental learning

## TL;DR

This paper introduces a framework for wearable devices that improves human activity recognition by balancing accuracy, speed, and power use in real-time.

## Contribution

A closed-loop adaptive learning framework that synergistically optimizes feature extraction, model sparsity, and online learning for edge devices.

## Key findings

- The framework achieves 85.6% to 97.4% accuracy across five datasets.
- Inference latency is approximately 1.0 ms, meeting real-time requirements.
- The system dynamically adapts to non-stationary sensor data streams.

## Abstract

Achieving high-precision, low-latency, and continuously adaptive human activity recognition on resource-constrained edge devices represents a core challenge. Existing research primarily focuses on improvements in single directions, such as “online learning,” “model sparsification,” or “feature extraction,” lacking a framework that synergistically optimizes all three. This leads to difficulties in dynamically balancing accuracy, latency, and power consumption when processing non-stationary sensor data streams.

To address this, this paper designs an end-to-end closed-loop adaptive learning framework. The core innovation of this framework lies in its system-level synergistic design: (1) Employing fast principal component analysis for adaptive feature dimensionality reduction; (2) Introducing an information theory-based dynamic sparse subnetwork activation mechanism to tackle the NP-hard problem of model selection; and (3) Integrating a low-complexity online incremental learning module for real-time tracking of concept drift. Through the closed-loop feedback and control of the aforementioned components, this framework achieves joint dynamic optimization of feature extraction, model complexity, and adaptation speed under edge computing constraints.

Experimental results across five datasets demonstrate that this framework achieves accuracies ranging from 85.6% to 97.4%, with inference latency of approximately 1.0 ms.

The framework comfortably meets the real-time requirement.