ATD: A Novel Statistical Framework for Sensor Fusion

University Of California-Santa Cruz, Santa Cruz CA

Investigators

Abstract

This project introduces a novel framework for the development of sensor fusion algorithms that incorporates ideas from the statistical literature on factor analysis. The project emphasizes applications to detection problems, and aims to develop flexible algorithms that are robust to violations of common assumptions such as Gaussianity of error distributions and linearity of transfer functions. The framework described by the investigators includes as special cases some of the most widely used tools for sensor fusion, such as linear and Kalman fusion filters, which are generalized to include the effects of non-linearities, non-Gaussian errors, concomitant variables, and correlations across sensors. One aspect that distinguishes this project from the more traditional literature on factor analysis (which has a long history, particularly in the social sciences) is that, in the context of sensor fusion, the latent factors have real physical meaning and therefore it is often possible to collect training sets that can be used to learn structural features of the model. The availability of these training sets allow the researchers to develop complex models for sensor fusion whose parameters would not be identifiable without them. In addition to providing a general framework for sensor fusion with wide applicability, this project also explores the application of these techniques to problems related to hyperspectral image analysis, particularly in the context of linear supervised and unsupervised unmixing. Recent technological advances have dramatically increase both the sources of data and the amount of data being collected in all kinds of fields. Making efficient use of these large amounts of information is a critical challenge in applications ranging from defense and national security to environmental sciences and industrial processes. This project develops the next generations of tools for sensor fusion, i.e., to optimally combine information arising from sensors located in multiple sites, or by monitoring a single at a very high frequency. The algorithms developed in this project will be more robust and generally applicable than most state-of-the-art approaches. In addition, because of the collaborations between the investigator and external groups at national laboratories and other government agencies, the tools developed in this proposal will have a deep impact on the ability of the Department of Defense to accomplish its missions.