Abstract
Nowadays, the prevalence of sensor networks has enabled tracking of thestates of dynamic objects for a wide spectrum of applications from autonomousdriving to environmental monitoring and urban planning. However, trackingreal-world objects often faces two key challenges: First, due to the limitationof individual sensors, state estimation needs to be solved in a collaborativeand distributed manner. Second, the objects' movement behavior is unknown, andneeds to be learned using sensor observations. In this work, for the firsttime, we formally formulate the problem of simultaneous state estimation andbehavior learning in a sensor network. We then propose a simple yet effectivesolution to this new problem by extending the Gaussian process-based Bayesfilters (GP-BayesFilters) to an online, distributed setting. The effectivenessof the proposed method is evaluated on tracking objects with unknown movementbehaviors using both synthetic data and data collected from a multi-robotplatform.