A single mobile device opportunistically trying to perceive the world around it is trying to solve an extremely difficult problem and has a limited perspective and limited computational resources with which to solve it. However, in many real-world situations there may be multiple mobile devices in the same proximity which are, at some level, all trying to solve essentially the same problem, maybe all for different end-goals. In principle, sharing data across these devices has the potential to boost perception accuracy by averaging out detection noise, circumventing occlusions, and making use of confident but unreliable sensors. In addition, sharing data in this fashion has the potential to increase the range of perception by orders-of-magnitude, allowing single devices to have an understanding not just of a room in which they are located, but also an entire floor, an entire building or even an entire city.
Federated Perception (FP) is a method for dynamically integrating probabilistic models across several users to infer events in space and time. Given a model specified for a single perspective plus a few simple rules, FP seeks to provide a mechanism to automatically and dynamically create large joint Bayesian models over many perspectives. The aim is to reduce the burden on the user for building complex joint models, and to build them ina way that can add or remove perspectives dynamically and that can simultaneously bound the compexity of the model. FP is also investigating methods to scale inference over many perspectives, including distributing computation across the individual users, and harnessing a cluster-computing backend. This dmo shows a proof-of-concept of FP in action for localizing and identifying objects in a cluttered environment.
The Federated Perception project is a subset of the Everyday Sensing and Perception (ESP) project and is strongly aligned with efforts in Computational Perception and the Data Rich Theme at Intel Research.