Bayesian Decision-Level Fusion Algorithm for Addressing Correlated Inputs

• Published in: Proceedings - IEEE Aerospace Conference pp. 1 - 11
• Main Authors: Agate, Craig S., Price, Jonathan D.
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.03.2025
• Subjects: Bayes methods; Classification algorithms; Correlation; Estimation; Fuses; Surveillance
• ISSN: 2996-2358
• DOI: 10.1109/AERO63441.2025.11068414

We consider the problem of correlated inputs to a Bayesian decision-level fusion algorithm. A decision-level fusion problem is characterized by combining the outputs of different classifiers/ATRs to more accurately and rapidly classify objects in a surveillance region. The outputs from the different classifiers may be declarations of a particular class type or may be a ranked ordering of class types based on confidence values. We presume that an individual classifier considers each look at an object separately from previous looks when arriving at declarations. Nonetheless, there may still be correlation in these subsequent declarations and ignoring this correlation in a Bayesian fusion rule can lead to inaccurate results. The correlation arises due to unknown parameters that affect the ATR algorithm's declaration (e.g., an object's pose in an image). In this paper, we focus on the situation in which a fusion algorithm sequentially fuses outputs from a single classifier. We describe the problem and highlight where the assumption of independent ATR outputs is typically made in the Bayesian fusion update. We then derive a joint estimation and classification algorithm that correctly addresses such correlation. The performance of the approach on a binary classification example with simulated measurements is given.