AcTrak - Unobtrusive activity detection and step counting using smartphones

Abstract

In this paper we introduce "AcTrak", a system that provides training-free and orientation-and-placement-independent step-counting and activity recognition on commercial mobile phones, using only 3D accelerometer. The proposed solution uses "step-frequency" as a feature to classify various activities. In order to filter out noise generated due to normal handling of the phone, while the user is otherwise physically stationary, AcTrak is armed with a novel algorithm for step validation termed as Individual Peak Analysis (IPA). IPA uses peak-height and inter-peak interval as features. AcTrak provides realtime step count. It also classifies current activity, and tags each activity with the associated steps, resulting in a detailed analysis of activity recognition. Using our model, a step-count accuracy of 98.9 % is achieved. Further, an accuracy of 95 % is achieved when classifying stationary, walking and running/jogging. When brisk-walking is added to the activity set, still a reasonable level of accuracy is achieved. Since AcTrak is largely orientation and position agnostic, and requires no prior training, this makes our approach truly ubiquitous. Classification of step-based activity is done as walking, brisk-walking and running (includes jogging). So, after a session of workout, the subject can easily self-assess his/her accomplishment.