Translator Disclaimer
25 April 2008 Example based learning of image stitching for an omni-directional camera using a variational optical flow methodology
Author Affiliations +
Omni-Directional vision plays an important role in autonomous and remotely controlled vehicles providing the critical ability of peripheral situational awareness. We introduce an omni-directional system which is able to build a high resolution uniform panoramic image from four different wide angled cameras. In order to build a uniform panoramic image, we developed a state of the art stitching algorithm using a variational optical flow estimation methodology. Optical flow is traditionally considered as the apparent 2D image motion captured by a single camera in different time samples. In this paper on the other hand, we consider optical flow as the 2D motion registering the overlap regions of images taken from different cameras at the same time instant. Since the rigid geometry between the cameras is fixed, the optical flow registering the different views is fixed for distant scenes. We use this fact in order to formulate a functional which requires that the same optical flow registers properly all the provided scene examples taken in the learning process. Our minimization functional incorporates in the data term all the available information as provided by the scene examples. We mathematically show that the variety of scene examples helps to overcome the aperture problem inherent in traditional optical flow problems. We demonstrate the robustness and accuracy of our method on synthetic test cases and on real images captured by our omni-directional commercial product.
© (2008) COPYRIGHT Society of Photo-Optical Instrumentation Engineers (SPIE). Downloading of the abstract is permitted for personal use only.
Tal Nir and Nir Karpel "Example based learning of image stitching for an omni-directional camera using a variational optical flow methodology", Proc. SPIE 7000, Optical and Digital Image Processing, 700024 (25 April 2008);

Back to Top