Real-time activity recognition in mobile phones based on its accelerometer data

Context awareness is one of the important keys in a pervasive and ubiquitous environment. Activity recognition by utilizing accelerometer sensor is one of the context aware studies that has attracted many researchers, even up until today. Inspired by these researches, we came out with this presented study, which is a continuation of our previous workswhere we explore the possibility of using accelerometer embedded in smartphones in recognizing basic user activity through client/server architecture. In this paper, we present our work in exploring the influence of training data size on recognition accuracy in building classifier model by studying two algorithms, Naïve Bayes and Instance Based classifier (IBk, k=3). The result shows that 13 out of 18 possible combinations for both algorithms gave 90% training data size as the best accuracy, thus proving the assumption that bigger size of training data gives better classification accuracy compared to small sized training data, in most cases. Based on the outcome from the study, it is then implemented in Actiware, which is an activity aware application prototype that uses built in accelerometer sensor in smartphones to perform real-time/online activity recognition. The recognition process is done by utilizing available phone resources locally, without the involvement of any external server connection. ActiWare manages to exhibit encouraging result by recognizing basic user activities with relatively small confusion when tested.