UBC Theses and Dissertations
Tracking the joints of articulated objects without an a priori shape model Leonard, Simon
The conventional methodology for tracking articulated objects relies on the knowledge of a shape model of the tracked object. The disadvantage of these top-down methods is their dependence on many assumptions, controlled environment and a priori knowledge. To be able to track and analyze arbitrary articulated motion, a radically different approach is necessary. This thesis proposes a method that addresses the lack of generality and adaptability of previous tracking systems for articulated objects. The method is built around an appropriate generic representation of articulated objects, which is a simple structural representation based on the joints configuration. The introduced method is a bottom-up approach able of detecting the joints of a moving object without any knowledge or assumptions. The method starts by extracting the moving contours of the object. Then the Hausdorff distance is used to decompose the contours into rigid components. From this decomposition, the joints are found by detecting the intersecting edge segments between adjacent components. Finally, the joints are then tracked by using an ad hoc deterministic version of the CONDENSATION algorithm.
Item Citations and Data