1 Feature-Based 3-D Surface Reconstruction Directed by Grid Space Projection Mohammed T. El-Sonni Wesam M. Dawoud College of Computing Engineering Arab Academy of Science and Technology and Maritime Transport P. O. BOX 1029, Alexandria, Egypt Abstract: - In this paper we describe a voxel-based 3-D reconstruction algorithm from multiple calibrated camera views. Unlike image-based algorithms, this algorithm is capable of detecting occlusion explicitly, and recovering the conventional Stereo Algorithms limitations; the algorithm is extendable to reconstruct the full surface without any restrictions on the cameras distribution. Because of using stable features at consistency checking the mismatching probability is decreased. The Grid Space is traced one time only; hence, there is no any additional computation or memory consuming. In spite of that, experimental results on both real and synthetic images show the algorithm efficiency and time preserving. Key-words:- Stereo Vision, 3-D reconstruction, Grid-Space, Feature-Based 1 Introduction Surface reconstruction algorithms have many practical applications from terrain mapping and industrial automation to Virtual Reality and real-time human- computer interaction. The 3-D reconstruction from multiple camera views can be distinguished into two classes of algorithms. The First class is the standard Stereo algorithms (image-based); they try to represent a scene from image pairs. A challenging is matching features between these images robustly. These methods typically employ normalized cross correlation along epipolar lines. Several authors have pursued volumetric representation to assist in this task, typically with image coordinates for two axis and disparity hypothesis being the third axis. The earliest example is given by [1] who develops a relaxation network that enforce uniqueness and continuity constraint by introducing inhibitory and excitatory connections between voxels representing disparity hypothesis. Marr at [1] uses discrete match values and a 2-D local support area possible due to memory and processing constraint, so his experiments was restricted to synthesized images, Zitnick [2] replace the 2-D local support area with 3-D local support area and construct a new update function, his algorithm can be run on both real and synthesized images. Zhang [3] extends the original cooperative algorithms [1] and [2] in two ways. First, he designed a method of adjusting the initial matching score volume to guarantee that correct matches have high matching scores. Second, he developed a scheme for choosing local support areas by enforcing the image segmentation information. As a result, the foreground fattening errors are drastically reduced. Numerous stereo matching algorithms, including local matching (e.g., [4], [5]), global optimization (e.g., [6], [7], [8], [9]), dynamic programming (e.g., [10],[11]) have been proposed over the past decades, all these methods are often limited for several reasons, First , input views can only be separated by a limited distance (baseline) , for correlation to be effective. Second, the result of stereo reconstruction is at best a 2½D reconstruction. Third, occlusion process is difficult to model image space. The Second class of 3-D reconstruction algorithms is object space (3-D space) algorithms, in 3-D space it is easer to reason