This paper addresses issues surrounding human activity recognition. While our main focus is health-care, the architecture presented in this paper is flexible enough to be applied to other domains. The architecture is the basis of a non-obtrusive system combining distributed and centralised paradigms founded on the concept of object networks. Object networks are created on the basis of user-to-object and object-to-object interactions. An object is consid-ered everyday artefact that is relevant in inferring user’s activity. The activity inference process is distributed throughout the environment and is carried out by the inference engine that operates over the object networks. This process actively exploits various events generated from the object network, describing the proximity relationships, or depicting specific object functionality in addi-tion to the state of the environment to infer user’s current activity. Moreover we utilise information representing user’s behaviour history as well as behav-iour of similar users in equivalent context to advance the activity inference process.