Contour Based Superquadric Tracking Jaka Krivic and Franc Solina Computer Vision Laboratory, Faculty of Computer and Information Science University of Ljubljana Trˇzaˇska 25, 1000 Ljubljana, Slovenia {jaka.krivic,franc.solina}@fri.uni-lj.si http://lrv.fri.uni-lj.si Abstract. This paper proposes a technique for tracking a superquadric- modelled object over a monocular video sequences. The object is cur- rently modelled with a single superquadric. Object’s position and orien- tation in the first frame of the sequence are assumed known. A frame in a sequence is first processed to find object’s contour. Contour is deter- mined by extracting edges on the frame in the vicinity of model’s contour from the previous frame. The model’s relative translation and rotation parameters are then calculated by fitting model’s contour to the frame’s contour. This fitting is achieved by minimizing the cost function, which is based on model to image mapping. Keywords: Superquadrics, object tracking, contours 1 Introduction and Motivation Object tracking from video sequences is a well researched area of computer vision, with a wide variety of possible applications. Common to these applications is that they need to extract information about object’s motion from image sequences. The model based approach to object tracking assumes that the object is somehow modelled, and the goal is then to match the model to the image. One of the more active fields of object tracking research is tracking of 3D articulated objects, especially human body, using 3D volumetric models [2, 3, 5, 6, 7, 8]. Tracking the object, while it is moving and articulating in front of the camera is done by fitting some kind of projection of the 3D model to the image. Superquadrics are a common building block for modelling articulated ob- jects [1]. One of the more appreciated properties is their very compact represen- tation, while on the other hand the failure to perfectly describe many natural shapes (especially the concave ones) is their main disadvantage [1]. Neverthe- less, in 3D human modelling they are quite frequently used [3, 7, 8, 4]. Although the models are not very photorealistic, they are very appropriate for high-level reasoning. Many human tracking methods use superquadrics for object modelling [3, 7, 8, 9]. In the approaches of Sminchisescu [7, 8], the superquadrics are used in for modelling only and are then turned to parameterized meshes. In our opinion, V. Palade, R.J. Howlett, and L.C. Jain (Eds.): KES 2003, LNAI 2774, pp. 1180–1186, 2003. c  Springer-Verlag Berlin Heidelberg 2003 brought to yo nd similar papers at core.ac.uk provid