Spectral Textile Detection in the VNIR/SWIR Band

Dismount detection, the detection of persons on the ground and outside of a vehicle, has applications in search and rescue, security, and surveillance. Spatial dismount detection methods lose e effectiveness at long ranges, and spectral dismount detection currently relies on detecting skin pixels. In scenarios where skin is not exposed, spectral textile detection is a more effective means of detecting dismounts. This thesis demonstrates the effectiveness of spectral textile detectors on both real and simulated hyperspectral remotely sensed data. Feature selection methods determine sets of wavebands relevant to spectral textile detection. Classifiers are trained on hyperspectral contact data with the selected wavebands, and classifier parameters are optimized to improve performance on a training set. Classifiers with optimized parameters are used to classify contact data with artificially added noise and remotely-sensed hyperspectral data. The performance of optimized classifiers on hyperspectral data is measured with Area Under the Curve (AUC) of the Receiver Operating Characteristic (ROC) curve. The best performances on the contact data are 0.892 and 0.872 for Multilayer Perceptrons (MLPs) and Support Vector Machines (SVMs), respectively. The best performances on the remotely-sensed data are AUC = 0.947 and AUC = 0.970 for MLPs and SVMs, respectively. The difference in classifier performance between the contact and remotely-sensed data is due to the greater variety of textiles represented in the contact data. Spectral textile detection is more reliable in scenarios with a small variety of textiles