JOURNAL OF INFORMATION SCIENCE AND ENGINEERING 24, 1719-1731 (2008) 1719 Estimating Optical Flow by Integrating Multi-Frame Information CHIA-MING WANG 1 , KUO-CHIN FAN 1,2 AND CHENG-TZU WANG 3 1 Institute of Computer Science and Information Engineering National Central University Chungli, 320 Taiwan 2 Department of Informatics Fo Guang University Ilan, 262 Taiwan 3 Institute of Computer Science National Taipei University of Education Taipei, 106 Taiwan In this paper, we develop two methods for estimating optical flow across multi- frames in an image sequence. The rationale of our proposed methods is based on inte- grating temporal information from the reference image and its previous and next several images. The difference between the two methods lies mainly on the modeling in utilizing multiple images. The proposed methods can be applied to each point independently, and hence is suitable for many image-based applications. Experiments were conducted on various image sequences and the results show that using multiple frames can indeed re- duce the estimation error significantly, especially at those points that possess good fea- tures. Keywords: multi-frame, optical flow estimation, gradient-based, first order approxima- tion, constant motion model 1. INTRODUCTION The task of motion estimation is to estimate the correspondence of moving objects in two or multiple frames in an image sequence which plays an important role in several computer vision applications, such as the recovering of 3D motion of objects relative to the viewer, the recognition of human faces and gaits, and the recovering of 3D structure of a scene or an object. Most of motion estimation approaches start from the computing of optical flow, which approximates two dimensional motion fields in the image domain. As we know, the computing of optical flow is a fundamental problem in the motion analysis of image sequences. Although Verri et al. [1] have proved that optical flow is in general different from the motion field − which is the projections of the motions of scene points relative to the observer − nevertheless, optical flow still provides significant mo- tion information which is a must to be used in many applications. Abundant of researches were conducted on the estimation of optical flow. Barron et al. [2] classified the optical flow algorithms into four categories including gradient-based, matching-based, energy- based, and phase-based approaches. Most of the optical flow algorithms are gradient- based and matching-based, while energy-based and phase-based approaches are not used frequently. The gradient-based methods, which was first proposed by Horn and Schunck [3], estimate optical flow depending on the image flow constancy equation deriving from Received February 16, 2007; revised May 27, 2007; accepted August 16, 2007. Communicated by Tong-Yee Lee.