Decentralised data fusion with particles
Lee‐Ling Sharon Ong, Ben Upcroft, Matthew Ridley, Tim Bailey, Salah Sukkarieh, Hugh F Durrant-Whyte · 2005
We aim to solve the problem of consistent Decentralised Data Fusion (DDF) with particle filters by a transformation of the sample statistics to a different representation that maintains an accurate summary of the particles. Two methodologies are proposed. The first method is a transformation of the particle representation to a Gaussian Mixture Model (GMM). The second algorithm approximates the particles by a Parzen representation. The two algorithms proposed differ in the accuracy of representing the particles as well as the accuracy of fusion methods and the bandwidth requirements. Our simulations results show that a transformation to GMMs requires less components and provides a more accurate summary compared to Parzen representations. However, the decentralised fusion solution for Parzen representations is more accurate than the solution for GMMs. 2. Communications are kept on a strictly node-to-node basis 3. There is no global knowledge of the network topology Practical applications of DDF have focused on representing features with Gaussian noise and through the use of ranging devices such as laser and sonar. While such techniques have been successfully used in autonomous air, ground, and underwater vehicles, constructing accurate models of unstructured and complex environments is difficult. However, our application aims to demonstrate DDF techniques for general non-Gaussian, non-point feature information. Such information includes observations of natural features and targets from both imaging and range sensors on flight and ground-based platforms such as in Figures 1 and 2. 1