Distributed particle filtering in agent networks: A survey, classification, and comparison
IEEE Signal Processing Magazine
D istributed particle filter (DPF) algorithms are sequential state estimation algorithms that are executed by a set of agents. Some or all of the agents perform local particle filtering and interact with other agents in order to calculate a global state estimate. DPF algorithms are attractive for large-scale, nonlinear, and non-Gaussian distributed estimation problems that often occur in applications involving agent networks (ANs). In this article, we present a survey, classification, and
... ison of various DPF approaches and algorithms available to date. Our emphasis is on decentralized ANs that do not include a central processing or control unit. Figure 1: Two examples of DPF applications in ANs: (a) target tracking, (b) chemical plume tracking. The agents are indicated by circles. the network or a relevant part thereof is an essential component of DPF algorithms. A standard application of DPFs in ANs is target tracking  (see Figure 1 (a)). A noncooperative target, such as a vehicle, aircraft, person, or animal moves through an area where the AN is deployed. The target emits a signal that is sensed by the agents, and a DPF estimates (tracks) time-varying properties of the target such as its position and velocity. A second application example is the tracking of chemical plumes  (see Figure 1 (b)). Here, micro aerial vehicles self-organize into an ad-hoc airborne AN and execute a DPF algorithm to estimate time-varying properties of the plume such as overall position, size, shape, and velocity. Distributed Estimation in Agent Networks Some examples of ANs are wireless sensor networks , sensor/actuator networks , robotic networks , networks of unmanned aerial vehicles (UAVs) , and networks of cameras . Possible applications of ANs include environmental and agricultural monitoring , health-care monitoring , target tracking , pollution source localization , chemical plume tracking , and surveillance . The agents may range from small, inexpensive, battery-powered sensor units equipped with limited computation and communication capabilities to resource-rich mobile robots or UAVs capable of performing complex tasks. Each agent contains (some of) the following elements: one or several sensors, communication interface, processing unit, and actuators. The on-board sensors measure physical quantities such as temperature, pressure, humidity, concentration of chemicals, distance, velocity, light intensity, vibration amplitude, or received signal power. The actuator elements allow the agents to act on the environment (e.g., turn on a water sprinkler to stop a fire) or on themselves (e.g., perform a controlled movement). In the application context considered in this article, the agents cooperatively estimate certain 2 parameters (or states) of the surrounding environment based on their local measurements. They need to cooperate because their local measurements are usually insufficient for obtaining reliable estimates. This is where DPF algorithms come into play. Compared to other distributed sequential estimation algorithms, DPF algorithms typically offer superior performance in nonlinear and non-Gaussian systems. For distributed estimation algorithms, communication aspects of the underlying AN are of central importance. These aspects concern the communication topology (which agents are connected by communication links) and the properties of the communication links (data rate, reliability, latency). The communication topology is commonly described by a communication graph, as briefly discussed in "Communication Graph of an Agent Network." For simplicity, we will usually assume the communication links to be error free. Communication Graph of an Agent Network The communication topology of an AN can be described by a communication graph G = (V, E) whose vertex set (or node set) V comprises the agents while each edge (k, k ′ ) ∈ E in the edge set E ⊆ V × V indicates an undirected communication link between agents k and k ′ . In the example shown in Figure 2 , the vertices are depicted by circles and the edges by lines connecting these circles. Figure 2: Example of a communication graph.