SBIR-STTR Award

Our proposal addresses the need for an algorithm to support the fusion of multiple and varied detector outputs into an actionable information. Specific focus is to identify and localize a radiation source in a complex environment and to characterize the complete scene. We will apply our extensive experience with radiation detection to thoroughly model a variety of radiation detectors deployed to complex scenes with multiple obstacles and various sources of radiation. Our proposed solution will be: 1) Robust – we use both statistical theory and recent developments in Robust AI for the US Government customers. Our machine learning methods are already being used with satellite data and applied to military space applications. Robustness of our methods will be thoroughly probed and proved through extensive testing and V&V activities. 2) Operationally efficient – achieved through detailed realistic modeling and extensive testing of the techniques we have recently developed for USAF customers. 3) Finely tuned to the specifics of the application – with the detailed realistic modeling of radiation detectors we develop algorithms that are specially designed to address requirements and constraints of the application. Our algorithms will enable quick and efficient decisions protecting the warriors in the field and contributing to the mission success.

This proposal for Phase II in response to the Defense Threat Reduction Agency (DTRA) topic 212-006 “Algorithm that can locally link radiation detectors (of different resolutions) to enhance identification/ localization capability”. The topic describes the need for an algorithm to support the fusion of multiple and varied detector outputs into actionable information. The specific focus is to identify and localize a radiation source in a complex environment and to characterize the complete scene. Our solution fully addresses these two goals. The approach that was taken by Verus Research in Phase I was to develop two classes of fusion algorithms using 1) traditional state estimation techniques and 2) modern unsupervised learning and supervised machine learning methods. These two approaches were compared using simulated, complex radiation scenes from a diverse set of radiation detectors. Input data have been provided by simulations performed with the use of simulation packages GEANT4, MCNP, GADRAS and RadSrc. Using these packages, we have obtained a large dataset of simulated detector measurements for various configurations of the radioactive sources. This dataset was used to train machine learning models and test the performance of all algorithms. Verus Research was able to demonstrate the development of a functioning algorithm capable of importing radiation detection data from multiple varied detectors, and fusing the varied outputs into actionable information. Our fused algorithm performance was proven to provide superior results compared to even best individual detectors included in this study. We plan to build on the success of Phase I to build a software product capable of using data from a variety of different detectors and integrate its results with existing platforms, such as TAK.