Modified SoftPOSIT algorithm for 3D visual tracking J. C. Diaz and M. Abderrahim RoboticsLab Universidad Carlos III de Madrid. Dpt. of Systems Engineering and Automation. Av. Universidad, 30. 28911. Leganes (Madrid), Spain. jcdiaz a c mohamed-*n Iuc3 es Abstract - This paper presents a new formulation of SoftPOSIT, common approach is to divide the search in two steps: first, a model based camera algorithm for determining the pose features are extracted from the image and second, a matching (position and orientation) of a 3D object from a single 2D image is looked for. Key factor of a pose estimation algorithm are when correspondences between object points and image points robustness with respect to noise in the input data and - are not known. This algorithm integrates the Softassign according to the application- speed. Input data are often technique for computing correspondences and the POSIT . . r c . . rf a technique for computing object pose. The method finds the nosisy tacted featuraecne ms poor fale .an their rotation and translation parameters of the camera with respect position may be inaccurate due to poor image quality, bad to an object. The major contribution of this work is in contriving light conditions, partial occlusions, precision of the a pragmatic approach for 3D pose estimation and tracking, acquisition and feature-extraction process. In addition, it is which yielded faster computation than the original algorithm worth noting that the pose problem actually consists of two and good target tracking performance. A new calculation of the sub problems: finding the position and orientation of the Distance Matrix which represents the relationship between the object and finding the correspondence between features. The features model projected and the image points has been pose problem implies finding the rotation and translation of introduced. This new approach has been successfully applied in the object with respect to the camera coordinate system, while synthetic and real images demonstrate the effectiveness of the the correspondence problem consists of establishing matching proposal modification. image features and model features. The problem is difficult Keywords - Model-based Pose Estimation, 3D object tracking, because it requires solution of two coupled problems, Modified SoftPOSIT. correspondence and pose, each one easy to solve only if the other has been solved first. Given matching between model and image features, one can determine the pose that best I. INTRODUCTION aligns those matches. If the object pose is known, one can simply determine such matches. Camera-based pose estimation of an real world object is an Most of the work presented in literature had addressed one important computer vision problem to solve because it has of the discussed sub-problems and recently there is clear influence upon several areas: object recognition, tracking, tendency to address them simultaneously as they appear naviatio a vt relt naturally in real problems. The 3D pose estimation problem inspection, autonomous has been approached with a variety of methods, like Neural [1,2,3,4,5]. Most of them are applied when scene models are Networks [8], linear programming [9] or Genetic Algorithms available, using visual landmarks for self-localization [6]. F This etails drawbck whih is te reconitionof natral [10]. For a full survey of different techniques of 3D object This entails a drawback whic is the recognimodeling correspondence and pose estimation [11] [12] and patterns. For this reason, usually artificial landmarks are g, p pc [ a i utilized to facilitate the computation [7]. However, artificial [13] ararecomended.correspondence and pose estimation landmarks involve a reorganization of environment which it is tatealywit the Scorrespondenc an poseheson noteve posibe. herfor, pse etemintio isa hrd simultaneously is SoftPOSIT [14]. This approach solve poteverposblem e Thercorespoden betweeminan 3D obje correspondence and pose problem matching 2D images and problem when the correspondences between 3D object *~~~~~apyn a deterministi annelin technique [6 and miieatvorepnimiznge 1Model-0basX0ed2.O ©200od Iae ueEo. oe fth beto