Robust and fast estimation of optical flow in computer vision

Yeon-Ho Kim, Purdue University


The apparent pixel motion in an image sequence, called optical flow, is a useful primitive for automatic scene analysis and various other applications of computer vision. Work done to date on the estimation of optical flow can be divided into two branches: increasing the accuracy for off-line applications, and reducing computational complexity for real-time applications. With regard to the accuracy issues, the two most common sources of error are: (1) the motion discontinuities induced by objects moving with respect to either other objects or with respect to the background; and (2) time varying illumination. Over the years, various robust estimation techniques have been proposed to surmount these twin challenges. These can be divided into methods that are global with respect to an entire image frame and methods that are local. Within the realm of local methods, previous researchers have proposed the least median of squares (LMS) as the most effective way to obtain a robust solution. This dissertation reviews this prior work and goes on to propose a modification of the LMS method that yields more accurate motion estimates at motion discontinuity boundaries. That brings us to the main contribution of this dissertation: An error analysis study that compares the standard LMS with our modification using a real image sequence. This error analysis study also includes a baseline implementation of LMS in which we use all of the available data, as opposed to a random sampling of the data, in the standard LMS and its modification proposed by us. Our error analysis demonstrates that, for the case of Gaussian noise, our modified LMS approach yields better estimates at moderate levels of noise. For the case of salt-and-pepper noise, the modified LMS method consistently performs better than the standard LMS method. In addition to these investigations into the accuracy-related issues pertaining to optical flow estimation, this dissertation also presents a fast voting-based approach for real-time optical-flow estimation. An additional topic addressed in this dissertation concerns stereo matching that uses estimated motions as an additional constraint in solving the correspondence problem. We will show some preliminary 3D reconstructions using this approach to stereo.




Kak, Purdue University.

Subject Area

Electrical engineering

Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server