Abstract. In this paper, we describe an approach to activity recognition, which is based on a self-organizing, ad hoc network of body-worn sensors. It makes best use of the available sensors, and autonomously adapts to dynamically varying sensor setups in terms of changing sensor availabilities, characteristics and on-body locations. For a widespread use of activity recognition systems, such an opportunistic approach is better suited than a fixed and application-specific deployment of sensor systems, as it unburdens the user from placing specific sensors at predefined locations on his body. The main contribution of this paper is the presentation of an interaction model for the self-organization of sensor nodes, which enables a cooperative recognition of activities according to the demands of a user's mobile device. We implemented it with an embedded system platform, and conducted an evaluation showing the feasibility and performance of our approach.