A generalized mean shift tracking algorithm  

作　　者：CHEN JianJun ZHANG SuoFei AN GuoCheng WU ZhenYang 

机构地区：[1]School of Information Science and Engineering, Southeast University, Nanjing 210096, China [2]Intelligence Engineering Lab, Institute of Software Chinese Academy of Sciences, Beijing 100190, China [3]Key Laboratory of Underwater Acoustic Signal Processing of Ministry of Education, Southeast University, Nanjing 210096, China [4]Nanjing Research Institute of Electronic Technology, Nanjing 210013, China

出　　处：《Science China(Information Sciences)》2011年第11期2373-2385,共13页中国科学（信息科学）（英文版）

基　　金：supported by the Open Research Foundation of Key Laboratory of Underwater Acoustic Signal Processing of Ministry of Education (Grant No. UASP1004);the National Natural Science Foundation of China(Grant No. 60672094);the National Basic Research Program of China (Grant No. 2009CB320804);the ChinaPostdoctoral Science Foundation (Grant No. 20100470588)

摘　　要：CAMSHIFT algorithm and Comaniciu/Meer algorithm are two fundamental frameworks of mean shift procedure for video target tracking. This paper generalizes the two well-known mean shift tracking al- gorithms, originally due to Bradski and Comaniciu/Meer. A new general similarity function which defines the distance between the target model and target candidate is employed to calculate the pixel weights and the target location. The target size is iteratively estimated and updated based on the zeroth order moment of the pixel weights. Then we prove that both the CAMSHIFT algorithm and the Comaniciu/Meer algorithm can be included in the generalized mean shift tracking framework. The tracking performances of three mean shift algorithms in the unified framework are shown and compared in the experimental results.CAMSHIFT algorithm and Comaniciu/Meer algorithm are two fundamental frameworks of mean shift procedure for video target tracking. This paper generalizes the two well-known mean shift tracking al- gorithms, originally due to Bradski and Comaniciu/Meer. A new general similarity function which defines the distance between the target model and target candidate is employed to calculate the pixel weights and the target location. The target size is iteratively estimated and updated based on the zeroth order moment of the pixel weights. Then we prove that both the CAMSHIFT algorithm and the Comaniciu/Meer algorithm can be included in the generalized mean shift tracking framework. The tracking performances of three mean shift algorithms in the unified framework are shown and compared in the experimental results.

关 键 词：mean shift CAMSHIFT video target tracking similarity measure 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] TN953[自动化与计算机技术—计算机科学与技术]