62 Chapter 5.  General Concepts, Toolkits, and Frameworks 5.1   State Of the Art Various research attempts have been made to develop frameworks and infrastructure for reusable sensing mechanisms. The context toolkit [ Salber et al. 1999 ] supports the development of context-aware applications with useful abstractions from the actual sensors. However, it mostly deals with context recognition on an abstract level decoupled from the variety of sensor technology. Furthermore, it limits applications to single sensor usage as only one context abstraction can be mapped to one physical sensor. In contrast to that, the TEA architecture [ Schmidt et al. 1999 ] focuses on low-level abstractions for simple sensors, which depends to much on the used sensors and, as such, does not provide reusable perception mechanisms either. The sensor classification scheme [ White 1987 ] facilitates the comparison and classification of sensors. Previously, several architectures have been proposed to model the flow of sensor data as it is being transformed to context and passed on to the application. The ContextToolkit [ Salber et al. 1999 ] separates the context acquisition process from the delivery and use of context. It borrows ideas from developments in graphical user interfaces by using widgets as a type of drivers for sensors.  Widgets encap- sulate states representing contextual information. Applications can be built upon these widgets either by polling or by registering call-back functions.  Beside wid- gets, context servers collect contextual data about a particular entity and context interpreters are used to transform context between di erent representations.  The context toolkit provides a solid infrastructure for managing context and querying context.  However, as the context toolkit uses widgets as sensor abstraction it does not give aid for integrating and querying the specific hardware sensors.  The issue of integrating di erent hardware sensors and perception mechanism on a single device is addressed in the TEA architecture [ Schmidt et al. 1999 ] .  This architecture defines layers focusing on the actual perception process, in particular on sensor fusion.  The distribution of components and the inclusion of domain knowledge in the perception process is not addressed. Context Fabric [ Hong and Landay 2001 ] is based on loosely coupled network ser- vices.  It defines an infrastructure for building context-aware applications based on a data store in which the context resides.  The main focus is on modelling, distribution and protection of contextual data.  The Location Stack [ Hightower et al. 2002 ] is an architecture specifically designed for combining di erent location sensing tech- niques.  It defines a common vocabulary for a set of multi-sensor location systems and implements probabilistic techniques for fusing the data.  The event heap devel- oped as part of the interactive workspaces project [ Johanson and Fox 2002 ] is based on a black-board approach.  Di erent parts of applications can use the Event Heap as a common message board to exchange data between loosely coupled components. The main focus of this project is on robustness and fault tolerance.  Applying this