A Novel Trust Region Tracking Algorithm Based on Kernel Density Estimation

Abstract

This paper presents a new approach which combines the Kernel Density Estimation and Trust Region algorithm for tracking objects in video sequences. Kernel density estimation (KDE) of the object's color distribution is built from the object region and used to generate a probability map for each incoming frame. Tracking is accomplished by localizing blobs in the maps. Compared with color histograms which are just empirical estimations of the objects' color distribution, KDE provides much better description of objects' color than histograms and promise better probability maps. The Trust Region algorithm ensures better convergence to objects' location than mean shift procedure. Different from the popular mean shift video tracking methods which determine objects' size and orientation using predefined parameters, the proposed algorithm calculates objects' size and orientation from geometric moments of the search window, rather than trial of discrete parameters. Experiments show that the proposed algorithm was able to precisely track the constant changes of the objects' size and orientation and achieved much better tracking precision on real video sequences than histogram based mean shift methods.