Obstacle Detection with Stereo Vision based on the Homography Nadia Baha and Slimane Larabi Computer Science Department, University of Science and Technology USTHB, Algiers, Algeria nbahatouzene@usthb.dz, slarabi@usthb.dz Abstract In this paper, we propose a simple method of obstacle detection that enables a mobile robot to locate obstacles in the indoor environment using an images pair from uncalibrated cameras. Using a set of features points that have been matched between the two views using the ZNCC correlation, a robust estimate of the homography of the ground is computed. The knowledge of this homography permits us to compensate for the motion of the ground and to detect obstacles as areas in the image that appear not stationary after the motion compensation. The resulting method does not require camera calibration, does not compute a dense disparity map and avoids the 3D reconstruction problem. This approach allows us to detect several numbers of obstacles of varied shapes and sizes. This obstacle detection stage can be viewed as the first stage of a free space estimator which can be implemented in an autonomous mobile robot. Keywords: Stereo images, Uncalibrated Camera, Obstacle detection, Homography. 1. Introduction In the last years, obstacles detection has been a major research topic in computer vision. However, it is still a hard problem to solve when automation, speed and precision are required and/or the objects present complex shapes. Indeed, obstacle avoidance is a fundamental requirement for autonomous mobile robots and vehicles, and numerous vision-based obstacle detection methods have been proposed. Some of them segment out obstacles from the ground plane based on differences of geometric properties, as the motion parallax [2,3,8,19], the projective invariant [12,13,16], and the depth information [1,20]. Others detect known obstacles based on their 2-D image pattern learned, beforehand [7]. Another technique common to mono and stereovision consist to use 3D scene reconstruction algorithms which are based on the triangulation operation [9,10,18]. However, a previous knowledge of the camera geometric model which is obtained by camera calibration is required. Most of the time, this step is delicate and sensible to measurement errors. In this paper, we propose a new approach of obstacle detection based on the image analysis from uncalibrated cameras for an autonomous mobile robot. The method assumes that the ground is planar and textured and starts by estimating the motion of the ground in the two images using a small set of matched points which permits us to make a robust estimate of the homography of the ground. Subsequently, compensation of the motion of the ground is performed by warping the second image with respect to the first according to the estimated motion. This warping registers the image of the ground in the two views, so that the obstacles are not stationary between the two images. Finally, a change detection operation between the first and the warped second image locates the obstacles present in the scene. Our main contribution in this work consists of the proposition of a simple method that does not require camera calibration, does not compute a dense disparity map, hence avoids the 3D reconstruction problem which are very time consuming. The advantage of this approach is that, it is very fast and permits us to detect a big number of obstacles of varied shapes and sizes. This paper is organized as follows: section 2 gives an overview of some preliminary results that are essential for the development of the proposed method. Section 3 presents the stages of the proposed method. We present in section 4, the experiments performed to demonstrate the feasibility and effectiveness of our approach. Finally, section 5 concludes the paper with some remarks. 2. First consideration 2.1. Projective Geometry In the following, projective (homogenous) coordinates are used to represent image points by 3 x 1 column vector m = (x,y , 1) T . An important concept in projective geometry is the plane homography H, which relates two uncalibrated views of plane in three dimensions. Each 3D plane P defines a non- singular 3 x 3 matrix H which links the image of the plane in two views. More specifically, if m is the projection in one view of a point belonging to the plan P and m’ is the corresponding projection in a second view, then [5]: m' H m (1)