Scalable Latent Parameter Estimation in Fusion Networks

Abstract:

Fusion networks aim to provide enhanced situation awareness in large scale defence & security applications in complex, dynamic environments. Together with other networked sensing applications involving spatio-temporal data streams, inference tasks based on the network-wide collected data (e.g., multi-object tracking) are underpinned by multi-sensor state space models. Estimation of latent parameters in these models underlies highly desirable capabilities such as sensor self-localisation in GPS denying environments, and, prediction of intent of vessels in surveillance. Conventional solutions, however, pose difficulties in scaling with the number of sensors due to the combinatorial complexity of joint multi-sensor filtering involved when evaluating the likelihood of the multi-sensor problem. We address this issue through a pseudo-likelihood approach and propose two families of surrogates for the parameter likelihood of the general multi-object tracking model. Owing to their node-wise separable structure, these approximations can be evaluated by local multi-object filtering operations, instead. When leveraged with pairwise Markov random field models and message passing algorithms for estimation (e.g., Belief propagation), they facilitate scalable estimation across the network and fit in both centralised and decentralised processing paradigms. The approximation quality is established in terms of how accurate the underlying state process is estimated locally. We demonstrate this approach for network self-calibration using measurements from non-cooperative targets in examples. We discuss future research directions in pseudo-likelihood design for approximate Bayesian inference motivated by these outcomes.

Biography:

Dr. Murat Uney is a Research Fellow in the School of Engineering, The University of Edinburgh. His research is in statistical signal and information processing with a particular interests on approximate and scalable statistical inference in distributed, multi-modal and resource constrained problem settings, and sensor fusion applications. Between 2010 and 2013 he was with Heriot-Watt University, where he developed multi-sensor fusion algorithms, and, demonstrated them online, on a maritime surveillance system in collaboration with BAE Systems and UCL. Prior to that he was a graduate researcher in the Signal Processing and Information Systems (SPIS) Laboratory, Sabancı University, İstanbul, and, a Ph.D. student at the Middle East Technical University (ODTÜ), Ankara. He was awarded his Ph.D. degree in electrical and electronics engineering with a major in signal processing and minor in control theory. He has industrial research and development experience both in the aerospace and telecommunications sectors. He is a member of the IEEE Signal Processing Society and the International Society of Information Fusion.