Deep Inference and Fusion Framework Utilizing Supporting Evidence (DIFFUSE)

Combining information from disparate sensors can lead to better situational awareness and improved inference performance; unfortunately, traditional multi-sensor fusion cannot capture complex dependencies among different objects in a scene, nor can it exploit context to further boost performance. Integrating context information within a fusion architecture to reason cohesively about scenes of interest has tremendous promise for refining the decision space, thereby improving decision accuracy, robustness, and efficiency. The Deep Inference and Fusion Framework Utilizing Supporting Evidence (DIFFUSE) program will produce a mathematical framework, founded on rigorous probabilistic analysis and learning theory, which will result in accurate modeling of information across different targets in the scene and context-dependent high-level semantic representation of target labels. In Phase I we will: (1) develop a probabilistic modeling and learning approach to model multi-sensor data and context-dependent target label correlations; (2) develop inference algorithms based on fused multi-sensor data and context for multi-target classification; and (3) demonstrate the implications of the proposed algorithms under various target classification scenarios. The results of the Phase I program will demonstrate the feasibility and promise of the DIFFUSE system concept to be realized in Phase II. Approved for Public Release 16-MDA-8620 (1 April 16)