ABSTRACT
Gkamas, Theodosios, N.
M.Sc., Computer Science Department,
University of Ioannina, Greece,
October, 2010.
Optical flow estimation using spatially varying smoothing.
Thesis’ Supervisor: Prof. Christophoros Nikou.
The problem of estimating the optical flow in a sequence of images is an important research problem in the area of computer vision with applications in visual object tracking, stereopsis and motion segmentation, among others.
Optical flow is the 2D velocity field, describing the apparent motion in the image that results from independently moving objects in the scene or from observer motion. Its estimation is a particularly difficult problem due to several factors.
At first, the massive image data which produce small and/or large scale linear systems that must be solved to obtain the solution in as little as possible and competitive period of time. Furthermore, the problems that occur because of the nature of the images, such as motion discontinuities and object occlusion must be addressed.
To overcome these difficulties, the majority of the state of the art optical flow computation techniques rely on the imposition of a smoothness constraint on the motion field.
In this work, we propose two methods for the accurate estimation of the optical flow where the smoothness constraint varies with respect to the image content. The first method is based on image segmentation and the smoothness constraint is applied to image areas belonging to the same segment and simultaneously presenting low spatial gradient information, to avoid smoothing probable motion boundaries. The second method relies on a probabilistic modeling of the optical flow problem where the motion vectors are considered as unobserved random variables generated by a Student’s t- distribution with spatially varying parameters. In that case, as the complete data likelihood is intractable we recur to the variational-Bayes methodology for inference of the model parameters and variables.
~ T. G. ~
Theodosios N. Gkamas, Ph.D. 