Visual contour tracking based on particle filters.pdf

61 Abstract—In the computer vision community, the Condensation algorithm has received considerable attention. Recently, it has been proven that the algorithm is one variant of particle filter (also known as sequential Monte Carlo filter, sequential importance sampling etc.). In sampling stage of Condensation, particles are drawn from the prior probability distribution of the state evolution transition, without making use of the most current observations, therefore, the algorithm demands a large number of particles and is computationally expensive. In this paper, a Kalman particle filter and an Unscented particle filter are presented to try to overcome the problem. These filters adopt sub-optimal proposal distributions, and use the Kalman filter or Unscented Kalman filter to incorporate the newest observation. This kind of sampling strategy can effectively steer the set of particles towards the region with high likelihood, and therefore, can considerably reduce the number of particles needed. Experiments with real image sequence are made to compare the performance of the three algorithms: Condensation, Kalman particle filter, and Unscented particle filter. Index Terms—Contour tracking, Kalman filter, Particle filter, Unscented Kalman filter, image sequences. I. INTRODUCTION ROBABILISTIC visual contour tracking has been an active research area in the computer vision community in the last ten or more years. It has many potential applications in intelligent robots, in monitoring and surveillance, in biomedical image analysis, in human-computer interfaces, etc. [1]. For these tracking tasks, a common approach is the use of the Kalman filter or extended Kalman filter. While some researchers employ physical snakes as system models in the (extended) Kalman filter [2,3,4], others use constant velocity motion models or learned motion models from training image sequence