Subtraction by Disparity Warping

In this thesis, I designed and implemented a background subtraction method based upon disparity verification that is invariant to run-time changes in illumination. Using two or more cameras, the method models the background by using disparity maps, which warp the primary image to each of the additional auxiliary images. During run-time, segmentation is performed by checking the color and luminosity values at corresponding pixels. When more than two cameras are available, segmentations become more robust and the occlusion shadows can generally be eliminated as well. The method assumes only fixed background geometry, which allows for illumination variation at run-time. And, because the three-dimensional reconstruction of the background is performed during the off-line stage, the method is easily implemented to achieve real-time performance on conventional hardware Thesis Supervisor: Aaron F. Bobick Title: Assistant Professor, MIT Media Laboratory