ParaLabel: Autonomous Parameter Learning for Cross-Domain Step Counting in Wearable Sensors

Abstract

Wearable step counters, also referred to as activity trackers, have been developed for health and activity monitoring, as well as for step tracking. These trackers, however, produce unreliable measurements during slow walking and when walking with assistive devices (i.e., aided walking). To address this challenge, in this article, we introduce, ParaLabel, a filter-based step counting algorithm that is reliable against various walking velocities and intensities. ParaLabel addresses this problem by learning a filter cut-off frequency autonomously in a new domain without the need for collecting sensor data and manually tuning the algorithm parameter for a different velocity and/or pattern of walking. We formulate this problem as a transfer learning problem in which the new filter cut-off frequency is transferred from a bank containing previously fine-tuned parameters from different domain(s). Our extensive analysis using real data collected from 15 participants while wearing an accelerometer sensor on their chest, wrist, or left pocket demonstrates the superiority of ParaLabel to two commercially available trackers worn on the same body location, and state-of-the-art techniques. ParaLabel achieves 96.3% - 99.9% accuracy during walking on a treadmill at three different velocities, 98.2% - 99.9% accuracy during walking with a shopping cart, and 89.3% - 97.3% accuracy while walking with the aid of a walker.