Comparison of color-based tracking by recursive filters Open Access
Downloadable ContentDownload PDF
Object tracking is required by many applications such as surveillance, human-computer interfaces, video communication or video compressing. The availability of high quality and the increasing demand of automated video analysis bring a great deal of interest in object tracking algorithm. In this paper, we provide an overview of tracking problems and the system model of this kind of problems. Since we are focusing on the video tracking, the characteristics of video tracking are analyzed in this project. Beginning with the Bayes' theorem, we presented the derivation of the tracking algorithms by Bayesian approach. We also discussed some assumptions that tracking problems often use. Based on these assumptions, we provide algorithms of Kalman filter, extended Kalman filter and particle filter. For each tracking filter, we summarize the parameters and equations we need to implement the algorithm.Then we go into details about a color-based tracking algorithm. Since our data containing target is a sequence of images, we use color distributions to represent the target. Color distribution is a robust method to describe color target. We also presented the method we used to evaluate the distance which is needed in tracking algorithm. Based on the above knowledge, we are able to provide detailed color-based Kalman filter, extended Kalman filter and particle filter algorithms.We implemented the tracking algorithms in some artificial videos first and tested the performance of the tracking filters. After that, we applied Kalman filter and particle filter on real video tracking problem. The tracking results are presented in the paper. After the result, we compared the performance of different tracking algorithms. Then we discussed the advantages and disadvantages of Kalman filter, extended Kalman filter and particle filter when applying to video tracking problems.