Hyperspectral band selection for human detection

Human detection based on spectral information is required for various applications, e.g. surveillance, tracking and missing person investigation. In practice, spectral human detection encounters the inherent challenge, i.e. multiple targets detection based on a limited number of spectral bands, because (1) there is a great variety in spectral profiles among various human-related materials, e.g. skin, natural fabric and chemical fabric, and (2) practical detection system will be implemented by multiband sensors rather than hyperspectral sensors. The objective in this paper is an optimal band selection for human detection by evaluating the distribution of targets and backgrounds in spectral space. The optimal combination of bands is selected by finding a hypersphere on which the target is placed in a maximum cavity surrounded by background distribution. The optimal band combination for multiple targets is selected by the maximization of the cavity size of the worst target.