| Real-time measurement of physical activity type on mobile devices could advance physical activity measurement studies and create new opportunities for real-time health interventions that promote physical activity. In this work, an algorithm for four-class activity detection (“Ambulation”, “Cycling”, “Sedentary” and “Other”) from wrist or ankle accelerometer-based activity monitors was extended to run in real-time on Android mobile phones with wireless accelerometers. Experiments validating the method using an existing activity dataset were replicated in the new Java-based system and results were verified to be consistent with the former MATLAB implementation by Mannini et al. The Android application was implemented with two modes: the continuous mode, which was suitable for instant classification (10s latency) and a power-efficient burst mode, which could be utilized in large scale study but with longer latency (up to 70s latency).;Although performance in the original work using MATLAB was good, early single-subject pilot testing of the real-time system exposed problems with the approach. Specifically, the models learned offline failed to learn important distinctions between some activities. For example for the data from ankle, some strenuous movements were classified as “Sedentary,” cycling with a desk cycle was sometimes classified as “Ambulation,” and running was classified as “Cycling.” To reduce errors, post-processing was applied to the classification results to capture “common sense” rules about overall motion and give activity prediction based on the data from both locations. After applying the post-processing methods, in single-subject pilot testing the modified real-time system classification improved with Wockets on either location.;In the continuous mode on a phone with a 1500 mAh battery, the system is able to provide real-time classification with voice prompt for approximately 8 hours; and in the power-efficient burst mode, the system operates for 24 hours. This work demonstrates the value of real-time activity recognition testing. |
