Articulated object tracking via a genetic algorithm

Within a human motion analysis system, body parts, are modeled by simple virtual 3D rigid objects. Its position and orientation parameters at frame t + 1 are estimated based on the parameters at frame t and the image intensity variation from frame t to t + 1, under kinematic constraints. A genetic algorithm calculates the 3D parameters that make a goal function that measures the intensity change minimum. The goal function is robust, so that outliers located especially near the virtual object projection borders have less effect on the estimation. Since the object's parameters are relative to the reference system, they are the same from different cameras, so more cameras are easily added, increasing the constraints over the same number of variables. Several successful experiments are presented for an arm motion and a leg motion from two and three cameras.