Texture based background subtraction
Dongxiang Zhou, Hong Zhang, Nilanjan Ray · 2008
Background subtraction is an effective technique for motion detection. A traditional background subtraction algorithm assumes a moving object (or objects) with respect to a static background, and segments the moving object(s) by classifying pixels into foreground and background with trained statistical models. Because classical background subtraction algorithms work with intensity images, they cannot handle situations in which all pixels are moving. To address this deficiency, we present a novel background subtraction algorithm in this paper that is capable of detecting objects of interest while all pixels are in motion. The key idea behind our algorithm is to work with feature images, rather than the raw intensity images, in which foreground and background exhibit sufficiently different statistics. We in particular use texture as the feature, extracted with circular Gabor filters at five different bands, to study the problem of detecting large objects (rocks) moving amid small fragments, in the application of detecting large frozen ore lumps traveling into a crusher. We will provide experimental results on real image sequences to illustrate the superior performance of our algorithm, compared with the classical intensity-based algorithm.