Cross-Platform Modelling for Human Activity Recognition System

Abstract: Human activity recognition (HAR) systems have a large set of potential applications in healthcare, e.g. fall detection and tracking physical activities. HAR systems based on wearable sensors have gained the most attraction, due to smartphones having these sensors embedded in them. This makes them a great candidate for collecting human activity sensor data. By utilizing the smartphone sensors, no other sensors need to be supplied and instead only a mobile application needs to be supplied. However, this comes with a trade-off, sensors embedded in smartphones display specific heterogeneity and biases, depending on platform and price range. Normally in such a scenario, multiple HAR systems have to be built and trained for each device. This is both a time consuming effort and gives no guarantees that the different systems will have similar activity recognition accuracy. Therefore, in this thesis, a HAR system is constructed, where classification methods and filtering techniques are explored and evaluated, in an effort to give some guidelines for how to construct a HAR system, that can be embedded in multiple platforms. This study shows that when considering a few common activities, this HAR system performs well even when sensor data is collected from multiple sources. Ensemble method AdaBoost, in combination with decision trees, gives the overall best performance. Filtering techniques, such as Butterworth and Chebyshev performs better than constant- and linear detrending. This is primarily due to their ability to distinguish between low frequency activities, such as standing and sitting. The best result in this study was given when combining Chebyshev filtering and AdaBoosted decision trees, with a F-score of 0.9877.