The research presented in this paper is a first investigation of the application of online unsupervised learning techniques in the case of smart buildings. An online learning approach is proposed to estimate the number of people (within a range) in a room based on data collected from standard sensors. This estimated range is changing and depends on the highest occupancy faced in a training dataset for instance. Occupant behavior is modeled using an online learning approach. The resulting algorithm makes use of recorded sensor data for motion detection, power consumption, and door position as well as acoustic pressure from a microphone. This paper presents an online algorithm for a mixture model-based clustering which has been applied to estimate the occupancy in an office context, with an excellent estimation accuracy.