Contextual information is comprised of a variety of different and heterogeneous sources of information. The dominant design approach for building context aware pervasive systems is a bottom up one. The crucial shortcoming of being of a bottom up nature is that the design of the system starts at the sensor layer. Subsequently a static, design-time configured, so called activity and context recognition chain is being build up on top of the sensor layer, used to infer contextual information out of the delivered sensor data. The definition of the system is performed during design time and is then kept static throughout its lifetime.
The incredible and irresistible rise of available smart gadgets with integrated sensing capabilities makes this approach antiquated. For a widespread use of context aware pervasive computing sensing ecosystems, new methodologies have to be discussed to revise the static and predefined nature of currently existing approaches.
The vision of this thesis is to carve the way to rethink and change the bottom up paradigm towards a goal oriented, dynamic, top-down configuration of a context aware system. A goal oriented methodology takes a so called recognition goal as input for a dynamic, self-organizing and adaptive system configuration during runtime. The goal oriented approach will revise the currently dominating methods and help to overcome the complexity crises of today's availability of trillions of sensing devices that can be used for Activity and Context recognition. Reducing the complexity of installing, configuring, optimising, and maintaining the sensing infrastructure in a goal oriented manner will help to make Activity and Context recognition systems successfully accepted and beneficial on a broader, open ended scale.
The goal oriented sensing approach follows an open world assumption, where sensing devices are assigned to goals according to their capabilities of contributing to the specified goals. A goal oriented sensing system can dynamically react and adapt to changes in the sensing ecosystem. This ensures, that at each point in time, the best selection of sensing entities is used according to the stated recognition goal.
The core contributions of this thesis are novel methodologies and algorithmic solutions to (i) define semantic Activity- and Context Relations, (ii) to formulate, translate and process Recognition Goals, that (iii) can be semantically matched to the available sensing infrastructure and dynamically configured during runtime, accompanied by (iv) making use of multiple sources of sensor information to reason the Activities and Contexts of the users.
The findings and contributions of this thesis are expected to induce a methodic shift away from predefined and static context aware systems towards their goal oriented and dynamic configuration and adaption during runtime based on a stated recognition goal.