Adaptive Threshold Estimation via Extreme Value Theory

Determining a detection threshold to automatically maintain a low false alarm rate is a challenging problem. In a number of different applications, the underlying parametric assumptions of most automatic target detection algorithms are invalid. Therefore, thresholds derived using these incorrect distribution assumptions do not produce desirable results when applied to real sensor data. Monte Carlo methods for threshold determination work well but tend to perform poorly when targets are present. In order to mitigate these effects, we propose an algorithm using extreme value theory through the use of the generalized Pareto distribution (GPD) and a Kolmogorov-Smirnov statistical test. Unlike previous work based on GPD estimates, this algorithm incorporates a way to adaptively maintain low false alarm rates in the presence of targets. Both synthetic and real-world detection results demonstrate the usefulness of this algorithm.