Distributed fusion with global and local variables

This paper presents a distributed fusion framework for probabilistic models with global and local variables, addressing scalable model aggregation in decentralized systems. The approach employs a conditional representation, expressing the probability density function (PDF) of global variables conditioned on local variables, supplemented by the marginal PDF of the local variables. By extending the logarithmic opinion pooling (logOP) rule to this representation, the framework enables efficient fusion of probabilistic models using practical distributed protocols such as flooding and consensus. It supports closed-form solutions for exponential family distributions and particle-based representations, leveraging a Rao-Blackwellized Particle Filter (RBPF) to marginalize local variables via particle filtering while maintaining analytical tractability for global variables. The framework is applied to distributed object tracking with multi-sensor registration. Simulations demonstrate near-centralized performance, validating the framework's scalability and effectiveness in fully decentralized environments.