Distributed Registration with Expectation-maximization and Multiresolution System for BMD

MRLets Technologies, Inc. proposes to develop a suite of algorithms and software for registration of track and report data from IR and radar sensors. The proposed registration system is able to jointly identify correspondence between data from distributed platforms and to estimate sensor bias parameters. The system relies on three technologies to meet the challenges of the project. First, a distributed multiple hypothesis registration is introduced to establish and evaluate global hypotheses of track/report correspondences. To this end, we developed 3D models for space-based IR sensors and ground-based X-band radar, from which the data from different platforms can be related to solve the true location of a target in a common coordinate system. The models enable the correlation of multiplatform sensor data, both metric and features, so that the likelihood of each hypothesis can be calculated. From the constraints on coexisting hypotheses, joint and marginal probabilities of all hypotheses can be estimated also, providing the most critical information for hypothesis maintenance and registration processing. In order to solve the problem of unresolved cluster and group targets, a spatial domain multiresolutional registration is introduced to build up multiresolutional track/report trees, from which the hypothesis of a correspondence can be made at a proper resolution. When sensor bias exists, a registration algorithm is required to estimate sensor bias parameters and subsequently remove the bias from sensor data. Through an expectation-maximization operation, we demonstrate that bias parameters can be estimated from maximizing the summation of log-likelihoods of registration hypotheses, weighted by their marginal probabilities. As expectation-maximization can be an iterative process, bias estimation and hypothesis processing can be jointly optimized when necessary. Besides removing one source of system error due to miss-matching of sensor resolution, registration over multiresolutional data makes it possible to estimate bias parameters more efficiently by iterating from a coarse resolution to a fine resolution. Overall, the proposed distributed registration with expectation-maximization and multiresolution system (DREAMS) provides a complete solution to the issues of sensor registration for a BMD application.