Item

Abstract

The rapid adoption of lightweight activity tracking sensors demonstrates that precise measures of physical activity hold great value for a wide variety of applications. The corresponding growth of physical activity data creates an urgent need for methods to integrate such data. In this paper, we demonstrate methods for 1) synchronizing accelerometer and Global Positioning System (GPS) data with optimal corrections for device-related time drift, and 2) producing principled estimates of step counts from GPS data. These methods improve the accuracy of time-resolved physical activity measures and permit pedestrian travel from either sensor to be expressed in terms of a common currency, step counts. We show that sensor-based estimates of step length correspond well with expectations based on independent measures, and functional relationships between step length, height, and movement speed expected from biomechanical models. Using 123 person-days of data in which Hadza hunter-gatherers wore both GPS devices and accelerometers, we find that GPS-based estimates of daily step counts have a good correspondence with accelerometer-recorded values. A multivariate linear model predicting daily step counts from distance walked, mean movement speed, and height has an R2 value of 0.96 and a mean absolute percent error of 16.8% (mean absolute error = 1,354 steps; mean steps per day = 15,800; n = 123). To best represent step count estimation error, we fit a Bayesian model and plot the distributions of step count estimates it generates. Our methods more accurately situate accelerometer-based measures of physical activity in space and time, and provide new avenues for comparative research in biomechanics and human movement ecology.