Purpose: Physical activity classifiers are typically trained on data obtained from sensors at a set orientation. Changes in this orientation (such as being on a different wrist) result in performance degradation. This work investigates a method to obtain sensor location and orientation invariance for classification of wrist-mounted accelerometry via a technique known as domain adaption. Methods: Data was gathered from 16 participants who wore accelerometers on both wrists. Physical activity classification models were created using data from each wrist and then used to predict activities when using data from the opposing wrist. Using subspace alignment domain adaption, this procedure was then repeated to align the training and testing data before the classification stage. Results: Prediction of activity when using data where the wearer’s wrist was incorrectly specified resulted in a significant (p = .01) decrease in performance of 12%. When using domain adaption this drop in performance became negligible (M difference < 1%, p = .73). Conclusion: Domain adaption is a valuable method for achieving accurate physical activity classification independent of sensor orientation in wrist-worn accelerometry.
Twaites and Hillsdon are with the Sport and Health Sciences Faculty; Everson is with the Computer Sciences Faculty; University of Exeter, Exeter, United Kingdom. Langford is with Activinsights Ltd., Kimbolton, United Kingdom.