Spectral Detection of Human Skin in VIS-SWIR Hyperspectral Imagery without Radiometric Calibration

Many spectral detection algorithms require precise ground truth measurements that are hand-selected in the image to apply radiometric calibration, converting image pixels into estimated reflectance vectors. That process is impractical for mobile, real-time hyperspectral target detection systems, which cannot empirically derive a pixel-to-reflectance relationship from objects in the image. Implementing automatic target recognition on high-speed snapshot hyperspectral cameras requires the ability to spectrally detect targets without performing radiometric calibration. This thesis demonstrates human skin detection on hyperspectral data collected at a high frame rate without using calibration panels, even as the illumination in the scene changes. Compared to an established skin detection method that requires calibration panels, the illumination-invariant methods in this thesis achieve nearly as good detection performance in sunny scenes and superior detection performance in cloudy scenes