Hyperspectral Based Skin Detection for Person of Interest Identification

An optimal skin identification system can have great impact in areas such as security and surveillance. In the field of security, skin identification can augment the effectiveness of other biometric security systems such as facial recognition and fingerprint identification [1]. In the case of surveillance, a real-time skin identification system will be extremely useful at tracking POI. Detecting and tracking using a skin identification technique provides positive, real time acquisition of people as they exit a building and prevents inadvertent track loss in a large crowd. This thesis presents the results of an ANN that is created in MATLAB® using the Neural Network Toolbox to identify a POI based on their skin spectral data. A baseline model is used with the optimal feature set identified by Cain [2]. The baseline model is then modified and optimized to maximize a ANN capability to identify a POI. Gaussian noise is calculated and added to the data sets to simulate atmospheric noise of a real world scene. The simulated atmospheric noise reduced the neural network\u27s accuracy by 14%. The neural network model is tested with real HSI images to verify the applicability of the ANN to identify a POI. The results for HSI testing are between 40-60% due to the illumination sources\u27 angle of incidence and the standard deviation of skin reflectance associated with differing skin locations on the body