A complete system for head tracking using Motion-Based Particle Filter and Randomly Perturbed Active Contour N. Bouaynaya and D. Schonfeld Department of Electrical and Computer Engineering, University of Illinois at Chicago, Chicago, USA. ABSTRACT Many real world applications in computer and multimedia such as augmented reality and environmental imaging require an elastic accurate contour around a tracked object. In the first part of the paper we introduce a novel tracking algorithm that combines a motion estimation technique with the Bayesian Importance Sampling framework. We use Adaptive Block Matching (ABM) as the motion estimation technique. We construct the proposal density from the estimated motion vector. The resulting algorithm requires a small number of particles for efficient tracking. The tracking is adaptive to different categories of motion even with a poor a priori knowledge of the system dynamics. Particulary off-line learning is not needed. A parametric representation of the object is used for tracking purposes. In the second part of the paper, we refine the tracking output from a parametric sample to an elastic contour around the object. We use a 1D active contour model based on a dynamic programming scheme to refine the output of the tracker. To improve the convergence of the active contour, we perform the optimization over a set of randomly perturbed initial conditions. Our experiments are applied to head tracking. We report promising tracking results in complex environments. Keywords: Video tracking, Particle filter, motion estimation, Adaptive Block Matching, active contour, snakes 1. INTRODUCTION Many real world applications require accurate people tracking. The traditional applications include video surveil- lance, 1 autonomous vehicle navigation, 2 human computer interfaces, robot localization, etc. Recent advances in computer, multimedia and communication technologies have created opportunities in new applications such as environmental imaging, 3 wireless communication 4 and virtual reality. 5 These emerging new applications pose new challenges and create an increasing need for developing more efficient and accurate tracking techniques. Recently Bayesian filtering framework has become very popular for object tracking. It provides a recursive formulation of the posterior probability density function in dynamical systems. Analytical solutions for the optimal Bayesian filtering problem are known only for special cases including the linear and gaussian case (Kalman filter 6 ). Particle filters provide a general framework for estimating the probability density function of general non-linear and non-Gaussian systems. They are based on a Monte Carlo approach, where the density is represented by a set of random samples. Samples can be drawn from any distribution called the proposal density or the importance function, but sample weights should be properly adjusted so that the sample set fairly approximates the posterior density. Theoretically, if the number of samples is sufficiently large, the sample approximation of the posterior density can be made arbitrarily accurate. 7 In practice, the samples have to be properly placed and weighted to get a fair approximation of the posterior distribution. The pioneering work of CONDENSATION 8 uses the state transition prior as its proposal distribution. Because the state transition does not take into account the most recent observations, the particles drawn from transition prior may have very low likelihood, and their contributions to the posterior estimation become negligible. Various improvements and extensions have been proposed for visual tracking, 9 , 10 . 11 To design better proposal distributions, the ICONDENSATION algorithm 9 uses a color tracker to generate the samples. This increases the sample set efficiency since it focuses the samples around the detected color blobs. However, this choice is ineffective in real world videos. The background clutter might have similar color properties as the object. Changing lighting conditions and/or object appearance causes the tracker to loose the object. In 11 an annealed Particle Filter has been proposed. It is based on probabilistic pruning and it focuses its particles around the global peaks of