Gradient-Enhanced Particle Filter for Vision-Based Motion Capture.pdfVIP

Gradient-Enhanced Particle Filter for Vision-Based Motion Capture Daniel Grest and Volker Kru?ger Aalborg University Copenhagen, Denmark Computer Vision and Machine Intelligence Lab {dag,vok}@cvmi.aau.dk Abstract. Tracking of rigid and articulated objects is usually addressed within a particle filter framework or by correspondence based gradient descent methods. We combine both methods, such that (a) the correspondence based estimation gains the advantage of the particle filter and becomes able to follow multiple hypotheses while (b) the particle filter becomes able to propagate the particles in a better manner and thus gets by with a smaller number of particles. Results on noisy synthetic depth data show that the new method is able to track motion correctly where the correspondence based method fails. Further experiments with real-world stereo data underline the advantages of our coupled method. 1 Introduction Motion tracking and human pose estimation are important applications in motion anal- ysis for sports and medical purposes. Motion capture products used in the film industry or for computer games are usually marker based to achieve high quality and fast pro- cessing. Marker-less motion capture approaches often rely on gradient based methods [13, 3, 19, 7, 10, 18]. These methods estimate the parameters of a human body model by minimizing differences between model and some kind of observations, e.g. depth data from stereo, visual hulls or silhouettes. Necessary for minimization are correspondences between model and observed data. The main problem of these correspondence based optimization methods is, that they often get stuck in wrong local minima. From this wrong estimated pose they can usually not recover. Other approaches like particle filters [5, 11] try to approximate the probability dis- tribution in the state space by a large number of particles (poses) and are therefore unlikely to get stuck in local minima, because they can follow and test a large num-