Modern building systems often utilize multiple, physically separated sensors (sometimes of different type) to better detect occupancy. Such systems typically rely on a central processor where global data fusion takes place. However, such centralized architectures give rise to problems with communication and computational bottlenecks and are susceptible to total system failure should the central facility fail. There are significant advantages in distributing operations over multiple processing nodes. In the wake of this need, this paper addresses the problem of presence detection in a building by employing a decentralized sensing architecture. We focus on a network of radar sensor nodes, each with its own processing facility. Each node runs a hidden Markov model algorithm to provide a local estimate of the occupancy state and share it via wireless links. We introduce a distributed fusion algorithm that optimally combines the information generated by local nodes, having access to their private information, and recovers exactly the global estimation. System performance is evaluated in real world conditions, where sensor errors and communication may not exactly follow idealized model assumptions.