You Are Sensing, but Are You Biased?

Abstract

Mobile devices are becoming pervasive to our daily lives: they follow us everywhere and we use them for much more than just communication. These devices are also equipped with a myriad of different sensors that have the potential to allow the tracking of human activities, user patterns, location, direction and much more. Following this direction, many movements including sports, quantified self, and mobile health ones are starting to heavily rely on this technology, making it pivotal that the sensors offer high accuracy. However, heterogeneity in hardware manufacturing, slight substrate differences, electronic interference as well as external disturbances are just few of the reasons that limit sensor output accuracy which in turn hinders sensor usage in applications which need very high granularity and precision, such as quantified-self applications. Although, calibration of sensors is a widely studied topic existing methods applicable to mobile devices not only require user interaction but they are also not adaptive to changes. Additionally, alternative approaches for performing more granular and accurate sensing exploit body- wide sensor networks using mobile phones and additional sensors; as one can imagine these techniques can be bulky, tedious and not particularly user friendly. Moreover, existing techniques for performing data corrections post-acquisition can produce inconsistent results as they miss important context from the device itself; which when used, has been shown to produce better results. In this paper we introduce a novel approach that exploits machine learning techniques to performan adaptive auto-calibration scheme for sensors with which achieves high output sensor accuracy when compared to state of the art techniques without requiring any user interaction or special equipment beyond device itself. Moreover, the energy costs associated with our approach are lower than the alternatives (such as Kalman filter based solutions) thus enabling our technique to be used efficiently on a wide variety of devices Finally, our evaluation illustrates that calibrated signals offer a tangible benefit in classification accuracy, ranging from 3 to 10%, over uncalibrated ones when using state of the art classifiers; we showthat for similar activities which are hard to distinguish otherwise, we reach an accuracy of > 95% where uncalibrated data classification only reaches 85%. This can be a make or break factor in the use of accelerometer data in health applications.