A Background Subtraction Algorithm for a Pan-tilt Camera

Abstract

This thesis is concerned with the detection of moving objects with a background subtraction algorithm from a pan-tilt camera. Traditionally, motion compensation is performed on the current image to align its pixels with their background models built temporally form their previous measurements. Appearance-based methods are sensitive to pixel misalignment during camera motion compensation. This problem can be alleviated by using pixel-wise motion such as optical flow, motion itself, can be inaccurate and, contributes to false positive foreground detection. In this thesis, we exploit the fact that pixel misalignment and inaccurate optical flow tend not to occur spatially simultaneously. Consequently, we can substantially improve the performance of the background subtraction algorithm by evaluating the marginal statistical models of appearance and motion separately – rather than jointly – in classifying whether a pixel is foreground. We conduct extensive experiments to validate our approach and establish its superiority to other competing algorithms in the literature.