Abstract
This paper investigates three background modelling
techniques that have potential to be robust against sudden and
gradual illumination changes for a single, stationary camera. The
first makes use of a modified local binary pattern that considers
both spatial texture and colour information. The second uses a
combination of a frame-based Gaussianity Test and a pixel-based
Shading Model to handle sudden illumination changes. The third
solution is an extension of a popular kernel density estimation
(KDE) technique from the temporal to spatio-temporal domain
using 9-dimensional data points instead of pixel intensity values
and a discrete hyperspherical kernel instead of a Gaussian kernel.
A number of experiments were performed to provide a com-
parison of these techniques in regard to classfication accuracy.