Abstract
The knowledge of the facing direction of a human can provide important behavioural cues that can be exploited for a diverse set of pervasive computing applications, ranging from improved human computer interactions and real world gaming to navigation and social signal processing. However, providing a machine understanding of the facing direction of humans continuously in different daily life situations remains a significant technical challenge, despite the increasing pervasiveness of computing technologies and advances in mobile computing. This thesis aims to close this gap by providing a solution for opportunistic estimation of the facing direction of smartphone users. While achieving this objective, it addresses a number of challenges faced by mobile phone-centric sensing systems including the volatile orientation and wearing position of the device in daily life situations and on-the-fly calibration of the sensors measurements against environmental impacts such as magnetic perturbation. Estimation of the facing direction of users is based on a novel technique, which exploits the acceleration pattern that can be measured by a smartphone as the user is walking. This approach is independent of the initial orientation of the device and is adaptable to various wearing positions on a user’s body, which gives the user a larger degree of freedom. The information about the wearing position is provided by an intelligent wearing position recognition technique, which constitutes our second contribution. Utilising a novel pre-processing approach alongside with a set of carefully selected features classifiers the proposed technique provides timely estimations of the wearing position of the device with almost perfect accuracy. Also a novel iterative algorithm is developed for fast and accurate calibration of the magnetometer readouts. The provided algorithm estimates and compensates the magnetic interference parameters with a minimum number of magnetic field samples while imposing a very low computation cost. Detailed evaluations of each individual technique against the existing state of the art techniques are provided. Throughout the evaluation procedures, prototypes of the techniques are developed on a smartphone when required and several field experiments are conducted. In order to prove the applicability of the provided approach for real world applications, two mobile phone application examples have been developed and evaluated in practice, which include a dead reckoning and a face-to-face interaction detection application. Our work provides an initial proof that conventional off-the-shelve smart phones can serve as a suitable platform for pervasive user direction estimation, provided that adequate algorithms are deployed on it.