3D tracking of deformable objects with applications to coding and recognition

In this contribution we address the problem of motion and structure estimation of objects that fit a deformation model. Our purpose is to provide a suitable input to a recognition system directed at detecting particular shapes and deformation patterns (gestures) of the object. This is accomplished by means of a stereoscopic vision system which first reconstructs 3D tokens-points-from the images; then the tokens are tracked independently in order to obtain an improved estimation of their positions and to keep a correspondence among them in consecutive instants of time. Finally the tokens are matched to an allowed state-shape-of a Finite State Machine which depicts the deformation of the body. Rigid motion is considered to relate the actual tokens positions with the estimated shape. This approach provides with a convenient way to deal with incomplete collections of measurements due to occlusions.