In Smart Home, understanding the environment and what is going on is the basis of all adapted services. Unfortunately, inferring situations and activity recognition directly from raw data is way too complex to be applied. Firstly, we present a layered architecture we are building to process raw data into abstract situations and activities. Secondly, data fusion tools using the belief functions theory are introduced as a general framework to provide a first level of abstraction from raw data given by sensors to a more complex context model. Then a methodology to apply the model to our Smart Home within the belief functions framework, a first implementation and the encountered issues in modeling are discussed.