Systems and Experiment Performance of Optical Flow Techniques.pdfVIP

International Journal of Computer Vision, 12:1, 43-77 (1994) ? 1994 Kluwer Academic Publishers, Manufactured in The Netherlands. Systems and Experiment Performance of Optical Flow Techniques J.L. BARRON Department of Computer Science, University of Western Ontario, London, Ontario, N6A 5tl7 D.J. FLEET Department of Computing Science, Queens University, Kingston, Ontario, K7L 3N6 S.S. BEAUCHEMIN Department of Computer Science, University of Western Ontario, London, Ontario, N6A 5B7 Abstract While different optical flow techniques continue to appear, there has been a lack of quantitative evaluation of exist- ing methods. For a common set of real and synthetic image sequences, we report the results of a number of regularly cited optical flow techniques, including instances of differential, matching, energy-based, and phase-based methods. Our comparisons are primarily empirical, and concentrate on the accuracy, reliability, and density of the velocity measurements; they show that performance can differ significantly among the techniques we implemented. 1 Introduction Without doubt, a fundamental problem in the process- ing of image sequences is the measurement of optical flow (or image velocity). The goal is to compute an approximation to the 2-D motion field--a projection of the 3-D velocities of surface points onto the imaging surface--from spatiotemporal patterns of image inten- sity (Horn 1986; Verri Poggio 1987). Once computed, the measurements of image velocity can be used for a wide variety of tasks ranging from passive scene inter- pretation to autonomous, active exploration. Of these, tasks such as the inference of egomotion and surface structure require that velocity measurements be accu- rate and dense, providing a close approximation to the 2-D motion field. Current techniques require that rela- tive errors in the optical flow be less than 10% (Barron et al. 1990; Jepson Heeger 1990). Verri and Poggio (1987) have suggested