Omnidirectional Depth Computation from a Single Image * Radu Orghidan † and El Mustapha Mouaddib ‡ ‡ Centre de Robotique Electrotechnique et d’Automatique Universit´ e de Picardie Jules Verne, Amiens, France. mouaddib@u-picardie.fr Joaquim Salvi † † Computer Vision and Robotics Group Institute of Informatics and Applications University of Girona, Girona, Catalonia, Spain. {radu,qsalvi}@eia.udg.es Abstract— Omnidirectional cameras offer a much wider field of view than the perspective ones and alleviate the problems due to occlusions. However, both types of cameras suffer from the lack of depth perception. A practical method for obtaining depth in computer vision is to project a known structured light pattern on the scene avoiding the problems and costs involved by stereo vision. This paper is focused on the idea of combining omnidirectional vision and structured light with the aim to provide 3D information about the scene. The resulting sensor is formed by a single catadioptric camera and an omnidirectional light projector. It is also discussed how this sensor can be used in robot navigation applications. Index Terms— catadioptrics, omnidirectional vision, cali- bration, structured light, 3D reconstruction I. I NTRODUCTION The omnidirectional vision sensors enhance the field of view of traditional cameras by means of special optics, structures of still or gyratory cameras or combinations of lenses and mirrors. Yagi [21] surveyed the existing techniques for building cameras with a wide field of view and Svoboda [19] proposed several classifications of the existing omnidirectional cameras according to their most important features. The catadioptric sensors use at least one mirror coupled to a conventional camera. The catadioptric cameras can be classified depending on the way they gather the light rays. When all the observed light rays cross into a point, called focus, the sensors are known as Single View Point (SVP). The class of catadioptric sensors that have a SVP was derived by Baker and Nayar [1]. The SVP is a desir- able property that enables distortion-free reconstruction of panoramic images in a familiar form for the human users. The catadioptric sensors that do not possess a single focal point (non-SVP) are less used but proved to be helpful for applications with specific requirements such as prescribed distortions [10] or with linear projection constraints [4]. Still, neither the standard cameras nor the catadioptric ones can provide depth information of the scene when used independently. Stereoscopic vision combines separate images taken from distinct points of view and permits to visually per- * This work is partially supported by the Spanish project CICYT TIC 2003-08106-C02-02 and by the AIRE mobility grant provided by the Generalitat of Catalunya that allowed a four month stay in the CREA lab from Amiens, France ceive depth. Stereo catadioptric sensors are special struc- tures of mirrors and lenses designed for obtaining depth from images with a wide field of view. In order to obtain distinct points of view of the scene the camera is pointed towards a structure of convex [13] [9] [2] or planar [8] mirrors. The 3D information is obtained by triangulation. This method leans on the assumption that the correspon- dences of the points between the observed images can be accurately found. However, correspondence matching is deteriorated in the case of catadioptric sensors because the resolution of the omnidirectional images is lower than the resolution of the conventional ones since the number of the scene points is significantly different while both images are represented using the same number of pixels. A solution to this problem is the use of a structured light pattern projected onto the scene [18] [17]. Using this technique is similar to placing visible landmarks in the scene so that image points can be identified and matched faster. It is noticeable that the use of 360 degrees images and of scene-depth information is ideal for robot navigation tasks. An efficient approach to the navigation problem is to choose a reasonable tradeoff between localization accuracy and travelled distance. Similarly, in real life a high precision of movements is required in small spaces such as narrow halls or offices while a low accuracy is needed in wide areas where the possible obstacles are more spaced. Consequently, robot navigation can be divided in two branches as shown by Gaspar [20]: topological navigation and visual path following. Topological navigation gives a qualitative characterization of the robot’s global position using omnidirectional images. This technique is suitable for navigation tasks that implies travelling large distances with low precision of localization. Visual path following is mainly used for short-distance segments that require high accuracy navigation. The goal of this paper is to present an omnidirectional sensor that provides 3D information using a single camera. From this point of view, a robot navigation application is also discussed. The sensor is formed by a single- camera catadioptric configuration with an embedded om- nidirectional structured light projector. By mounting the omnidirectional sensor on a mobile robot applications such as 3D map building, robot navigation and localization, active surveillance with real-time object detection or 3D Proceedings of the 2005 IEEE International Conference on Robotics and Automation Barcelona, Spain, April 2005 0-7803-8914-X/05/$20.00 ©2005 IEEE. 1222 Authorized licensed use limited to: UNIVERSITAT DE GIRONA. Downloaded on April 26,2010 at 10:44:08 UTC from IEEE Xplore. Restrictions apply.