Currently it is difficult to determine when and where people occupy a commercial building. Part of the difficulty arises from shortcomings in available sensor technology, but an even greater deficiency is the lack of analysis methods appropriate to the determination of occupancy. This paper describes a pilot study describing new sensing and data analysis techniques, applied to the determination of space occupancy. The central premise of the paper is that improved building operation with respect to energy management, security, and indoor environmental quality will be possible with better detection of building occupancy resolved in space and time. We developed and deployed a network of passive infrared occupancy sensors in two private offices, and applied analysis tools based on Bayesian probability theory to determine occupancy. Specifically, a class of graphical probability models, called belief networks, was applied to the occupancy data generated by the sensor network. The inference of primary importance is a probability distribution over the number of occupants and their locations in a building, given past and present sensor measurements. Inferences were computed for occupancy and its temporal persistence in individual offices as well as the persistence of sensor status. The raw sensor data were also used to calibrate the sensor belief network, including the occupancy transition matrix used in the Markov model. sensor sensitivity, and sensor failure models. This study shows that the belief network framework can be applied to the analysis of data streams from sensor networks, offering significant benefits to building operation compared to current practice. (C) 2006 Elsevier B.V. All fights reserved.