Distributed particle filtering in agent networks: A survey, classification, and comparison]

Distributed particle filter (DPFs) are a powerful and versatile approach to decentralized state estimation in agent networks (ANs), and they are especially suited to large-scale, nonlinear, and non-Gaussian systems. Most distributed discrete-time sequential estimation algorithms presuppose synchronization, the availability of a common clock or time base at each agent, whereas some approaches relax this requirement. The existing DPF algorithms differ in aspects such as type and amount of the data communicated between the agents, communication range, local processing, computational complexity, memory requirements, estimation accuracy, robustness, scalability, and latency. In fusion center (FC) based DPFs, each agent uses a local PF to convert its own measurement into a local posterior, which is then transmitted to an FC. In leader agent (LA) DPFs information accumulates along a path formed by a sequence of adjacent agents. With consensus-based DPFs, all agents perform particle filtering simultaneously and possess a particle representation of a posterior.