HIERARCHICAL BLOCK-BASED DISPARITY ESTIMATION USING MEAN ABSOLUTE DIFFERENCE AND DYNAMIC PR.pdf

HIERARCHICAL BLOCK-BASED DISPARITY ESTIMATION USING MEAN ABSOLUTE DIFFERENCE AND DYNAMIC PROGRAMMING Liang Zhang Communications Research Centre Canada 3701 Carling Avenue, Ottawa, Ontario, K2H 8S2, Canada email: liang.zhang@crc.ca, Fax: (613) 990-6488, Tel: (613) 991-6995 ABSTRACT A disparity field is required in many video processing and computer vision applications. Many current approaches to disparity estimation tend to fail in low textured areas unless an additional interpolation step is added. To overcome this problem, an approach based on the hierarchical block-matching technique is proposed. To avoid ambiguous matching problems in low textured areas, a cost function is developed, which considers similarity measure of current corresponding points and estimates of neighboring disparities. For similarity measure, the mean absolute difference of rectangular blocks surrounding the points to be evaluated is exploited. Cost minimization is carried out by means of dynamic programming along each scan line. Preliminary results show that the percent of matching points is increased by 4% and computation time is reduced by 70% compared to the current approach. 1. INTRODUCTION Estimation of the disparity field in natural image pairs plays an important role in many video processing and computer vision applications, such as multi-viewpoint image generation [4][6][8][10]. For estimating the disparity field, many approaches have been proposed. These can be divided into three classes: feature-based, pixel-based and area-based approach. The feature-based approach extracts features from image pairs and then matches these features [9]. The pixel- based approach [1] finds corresponding points by measuring individual pixel intensity difference between image pairs and by considering some geometric constraints during cost minimization. With this approach, disparity values in low textured areas can be derived. The area-based approach [2][4][7] finds corresponding points by measuri