Background Image Subtraction Using Multi-view Geometry for Wide Baseline Image

Background subtraction has found its application for tracking and 3D reconstruction. The challenge remains in identifying a foreground object from a given set of 2D images. Such a framework needs to take into account shadows, occlusion, intensity variations and movements in the background. This project proposes one such background subtraction method which makes use of multi-view geometry. While such methods exist till date for stereo images, little research has been done for wide-baseline images. A novel approach to performing background subtraction across such wide-baseline images is proposed. For the purposes of background subtraction a dense pixel-to-pixel correspondence map is generated across three wide-baseline images. Using this correspondence data, a background subtraction method, which is invariant to shadows and intensity variation is proposed. The method is tested over wide-baseline example set with single and multiple-persons as foreground objects. A thorough evaluation of the results is done using epipolar geometry. Promising results are obtained even at 20 frames per second. The obtained results are compared with the results using the " Kernel Density Estimation " method and the conventional frame differencing to test its quality. Several extensions to the approach are proposed to further improve the quality of the foreground images including an extension to N-view images. Acknowledgements My sincere thanks to the entire team at Utrecht University working on this project, Nico van der Aa, Robby Tan, Xinghan Luo and Remco Veltkamp. I would like to thank Nico van der Aa for his invaluable suggestions and feedback on the project and continuous support throughout the phase of the project. Without his continued persistence and guidance I would not have managed to finish this project in the stipulated time-frame. Sincere thanks to Remco Veltkamp for providing an opportunity to work in such a challenging topic. Thanks to Robby Tan for his feedback on concepts of computer vision which were used throughout the progress of the project. Very special thanks to Xinghan for providing the calibration data and the input dataset.