Computing context is a major subject of interest in smart homes. In this paper, we present how we adapted a general purpose multi-level architecture for the computation of contextual data to a prototype of smart home. After a quick explanation of why we use different methods at different levels of abstraction, we focus more on the low-level data fusion. To do this, we present the basics of belief functions theory and how we apply this theory to sensors to obtain stable abstractions. By doing this, we highlight the major problem appearing when processing directly sensor measures. We respond to this problem by introducing an abstraction of sensors we call virtual sensors. Some examples of virtual sensors are given.