Illumination invariant foreground detection algorithms for real time applications

Abstract

In this thesis new foreground detection techniques are proposed to detect moving objects when various illumination changes occur in the monitoring scene Illumination variation is an important problem in many computer vision applications The objective of this research is to design algorithms that are efficient in foreground detection and can run in real time Foreground detection is a primary step in many computer vision applications It mostly employs background subtraction algorithms to detect moving objects in monitoring scenes because of its simplicity and efficiency Initially a thorough study on various state of the art background subtraction algorithms was performed to access its ability to efficiently extract the foreground region when the monitoring scene is influenced by illumination changes To overcome the disadvantages of existing algorithms four new background subtraction techniques are proposed to extract foreground objects in real time These four techniques can be divided into two groups by their way of illumination change detection The first group of algorithms uses entropy as a measure and the second group of algorithms uses luminance and structural similarity as a measure to detect changes in the video sequence In the first two proposed algorithms the entropy of the scene is estimated to detect illumination changes in the monitoring area The first proposed method replaces the current background model of the scene by an ideal background model whenever extreme illumination changes are detected but in the second algorithm the background model is updated at different learning rates by classifying the illumination changes into three levels newline