Learning a Physical Activity Classifier for a Low-power Embedded Wrist-located Device

TL;DR

This paper introduces a new real-time activity classification algorithm optimized for low-power wrist devices, demonstrating high accuracy and efficiency on commercial System-on-Chips with extensive user data.

Contribution

The paper presents a novel activity classifier algorithm tailored for low-power embedded wrist devices, optimized for real-time operation and evaluated on a large user database.

Findings

Achieved 96% accuracy for Rest, 94% for Walk, and 99% for Run.

System generalizes well to activities like XC skiing and Housework.

Benchmarking shows low memory footprint and power consumption.

Abstract

This article presents and evaluates a novel algorithm for learning a physical activity classifier for a low-power embedded wrist-located device. The overall system is designed for real-time execution and it is implemented in the commercial low-power System-on-Chips nRF51 and nRF52. Results were obtained using a database composed of 140 users containing more than 340 hours of labeled raw acceleration data. The final precision achieved for the most important classes, (Rest, Walk, and Run), was of 96%, 94%, and 99% and it generalizes to compound activities such as XC skiing or Housework. We conclude with a benchmarking of the system in terms of memory footprint and power consumption.