Yong Rui

报告题目: Multicue Energy-Driven HMM-UKF for Real-Time Object Tracking
 
报告人: Dr. Yong Rui
              Microsoft Research Redmond, USA
              Associate Editor of IEEE Trans. Multimedia
              Editor of ACM/Springer Multimedia Systems Journal
              Program Chair of ACM MM'06, CIVR'06

报告时间: 2005年10月18日15:00 - 16:00

报告地点：蒙民伟楼404会议室 
 
摘要:
Two difficulties in object tracking hinder its real-world usage: how to incorporate multiple visual cues and how to handle the nonlinearity in the system efficiently and robustly. To address these two issues, we present a new tracking method combining hidden Markov model (HMM) and unscented Kalman filter (UKF). First, based on a parametric shape model, an HMM is designed to detect the object contour based on multiple cues and contour smoothness constraint. To handle background clutter, we further utilize the data association technique to track the multiple edges (i.e. the true contour and background clutter) jointly. Finally, system dynamics and the probabilistic contour detection results from HMM are combined by a UKF to estimate the object states. Better than other variants of the recursive least mean square estimators, UKF approximates nonlinear systems up to the second order. The proposed HMM-UKF paradigm provides a powerful energy driven solution for multicue nonlinear object tracking. Further model adaptation to changing environments can be achieved by the Balm-Welch algorithm. The new algorithm achieves promising results in various complex real-world video.