Comparing human target detection with multidimensional matched filtering methods

Abstract

Recent technological advances in sensor manufacturing enable the use of separate spectral bands; e.g., MWIR and LWIR, to generate spatially registered imagery. Human factors experiments can be used to test whether a sensor can improve operator performance for detecting or recognizing a target. Although human factors experiments are of tremendous value, these tests are time consuming and resource intensive. In order to reduce costs associated with collecting behavioral data, an alternative approach is discussed. We propose using signal detection theory, to compliment and reduce the amount of classical human performance testing. As a test case we have studied whether multi-spectral sensors are significantly better than single band sensors. Scribner, Satyshur, and Kruer (1993) demonstrated that a two­ dimensional matched filter (spatial) optimized for a specific target and background power spectra, can be used to estimate an observer's ability to detect the target embedded in a cluttered background. Three different background images were used with, and without, a target present. False alarm and target detection probabilities were computed and results were plotted on a Receiver Operating Characteristic (ROC) curve. The matched filter ROC curves were then compared to behavioral ROC curves. Results showed that the matched filter ROC curves were similar to behavioral ROC curves with color fusion and long-wave infrared showing the highest sensitivity and mid-wave and short-wave infrared scenes were significantly less sensitive. These results indicate that the matched filter analysis may be used to model human behavior.