Robust Detection of Radiation Threat

Background. We consider the problem of radiological source detection in the presence of noisy background environments. The objective is to detect dangerous materials sources by their unique radioactive signatures. The foundations of this problem are presented here. Related work. There has been much research on statistical and signal processing techniques for source detection. These methods mostly concentrate on anomaly detection when one has little information about the target source. The most popular class of methods involves Principal Components Analysis to learn the expected characteristics of background. There are also many detection methods that do not involve Principal Components Analysis. Existing methods. We describe PCA anomaly detection, Censored Energy Window, Matched Filter, and Gaussian-Poisson MAP estimation in detail. Uncertain Censored Energy Windows. A method is proposed for a setting in which information about threat templates is incomplete. Source independence experiment. An experiment is performed that simulates incomplete knowledge of the source template. Simultaneous estimation of source strength and background. Another method is proposed that has much lesser dependence on training data. Generalization experiment. An experiment is conducted by simulating test data that greatly differ from training data. Conclusion. Both new methods are empirically demonstrated to perform better than competitors.