Visual Object Tracking Based On Realtime Evolutionary Particle Filters

Hoon Kang: Intelligent Robot & Vision Lab., School of Electrical and Electronics Engineering, Chung-Ang University, 221, Heukseok-Dong, Dongjak-Gu, Seoul, 156-756, KOREA Tel : +82-2-820-5320 Fax : +82-2-816-1856 Mobile Phone : +82-11-352-4810 E-mail : hkang@cau.ac.kr Homepage URL : http://sirius.cau.ac.kr

Abstract

We investigate several issues of visual tracking of multiple objects. First, we address a visual object tracking algorithm based upon particle filters (PF) which combine multiple observation models such as the active contour and the point-cloud measurements of the interframe absolute difference (IFAD) features and/or the target HSV colors. The former is applied to matching the state-affine-deformable contour of the moving targets and the latter is used to independently enhance the likelihood of tracking with a particular target color-cluster. Particle filters are more efficient than any other tracking algorithms because the tracking mechanism follows a bayesian inference rule of conditional posterior propability density propagation which resembles both the selection and the mutation operators of the evolutionary strategy (ES) even if it suffer from the impoverishment issue. Second, we upgraded the tracking performance by applying the competitive-AVQ neural network to resolving reinitialization problems when new objects appear to the camera viewport. Moreover, we solved overlapping or occlusion problems of multiple moving targets by using the priority-based subtractive potential fields of the IFAD features. Third, we found that only IFAD features can still be detected even when tracking moving objects with the pan-tilt moving camera. Fourth, we also developed a segmentation tracking algorithm by using the angle-sorted circular-checking routine to discriminate the multiple foregrounds from the background. Finally, we demonstrated the possibility of real-time object tracking as an intelligent interface by simulating the deformable contour particle filters. In the experimental results, it is shown that the suggested contour tracking particle filters prove to be robust under the light changes and in the cluttered environment.

Keywords: Visual Tracking, Robot Vision, Evolutionary Strategy, Particle Filter, Adaptive Vector Quantization, Conditional Posterior Probability Density Propagation, Active Contour, Segmentation

Short Biography

Hoon Kang: He was born in Seoul, Korea, in 1959. He received the B.S. and M.S. degrees in electronics engineering from the Seoul National University, Korea, in 1982 and 1984, respectively. He earned Ph.D. degree and the CIMS certificate in electrical engineering from Georgia Institute of Technology, Atlanta, USA, in 1989. From 1989 to 1991 he was first a postdoctoral fellow and then a research associate in the Georgia Tech Electrical Engineering Department and MARC. As he participated in a number of projects sponsored by the Office of Naval Research, the Ford Motor Company, and Honeywell, he developed new ideas on fuzzy logic control, intelligent robotics, and fault detection and identification. Since 1992, he joined the department of control and instrumentation engineering, Chung Ang University, Seoul, Korea as an assistant professor. He is now a professor at the school of electrical and electronics engineering, Chung-Ang University. He is a member of IEEE, IEEK, ICASE, and KFIS, where he has served as a number of committee members. His interests are robot vision, visual tracking and recognition, computational intelligence (FS, NN, GA, GP, ES, EP), artificial intelligence, and artificial life.