SCHOOL OF ENGINEERING - STI ELECTRICAL ENGINEERING INSTITUTE SIGNAL PROCESSING LABORATORIES Pablo Revuelta Sanz EPFL - STI - IEL - LTS5 Station 11 CH-1015 LAUSANNE Phone: +41 21 693 78 91 E-mail address: prevuelt@ing.uc3m.es Reference: EPFL-REPORT-150511 Stereo Vision Matching using Characteristics Vectors Pablo Revuelta Sanz 1 1 Carlos III University of Madrid, Spanish Center for Captioning and Audiodescription. , Belén Ruiz Mezcua 1 , José M. Sánchez Pena 1 , Jean-Phillippe Thiran 2 Av. Peces Barba, 1. 28918 Leganés, Madrid. Spain. 2 Signal Processing Laboratory (LTS5), École Polytechnique Fédéral de Lausanne. (EPFL), Station 11, CH1015 Lausanne, Switzerland. Abstract Stereo vision is a usual method to obtain depth information from images. The problems encountered when applying the majority of well established algorithms to provide this information are due to the high computational load required. This occurs in both the block matching and graphical cues (such as edges) matching. In this article we address this issue by performing an image analysis which considers each pixel only once, thus enhancing the efficiency of the image processing. Additionally, when matching is carried out over statistical descriptors of the image regions, commonly referred to as characteristic vectors, whose number of these vectors is, by definition, lower than the possible block matching possibilities, the algorithm achieves an improved level of performance. In this paper we present a new algorithm which has been specifically designed to solve the commonly observed problems which arise from other well know techniques. This algorithm was designed using a previous work carried out by the authors in this area to determine the descriptors extraction processes. The complete analysis has been carried out over gray scale images. The results obtained from both real and synthetic images are presented in terms