Distributed signal processing for wireless sensor networks enables thatdifferent devices cooperate to solve different signal processing tasks. Acrucial first step is to answer the question: who observes what? Recently,several distributed algorithms have been proposed, which frame thesignal/object labelling problem in terms of cluster analysis after extractingsource-specific features, however, the number of clusters is assumed to beknown. We propose a new method called Gravitational Clustering (GC) toadaptively estimate the time-varying number of clusters based on a set offeature vectors. The key idea is to exploit the physical principle ofgravitational force between mass units: streaming-in feature vectors areconsidered as mass units of fixed position in the feature space, around whichmobile mass units are injected at each time instant. The cluster enumerationexploits the fact that the highest attraction on the mobile mass units isexerted by regions with a high density of feature vectors, i.e., gravitationalclusters. By sharing estimates among neighboring nodes via adiffusion-adaptation scheme, cooperative and distributed cluster enumeration isachieved. Numerical experiments concerning robustness against outliers,convergence and computational complexity are conducted. The application in adistributed cooperative multi-view camera network illustrates the applicabilityto real-world problems.
展开▼
