Adaboost Blob Tracking
Jingping Jia · 2012
This paper presents a new approach for tracking blobs in image sequences in which tracking is seen as a binary classification problem. Firstly, the linear combination of R, G, and B with integer coefficients are used to build weak classifiers. Then the Adaboost learning schema is employed to construct a strong classifier from those weak classifiers which have large two-class variance ratio. For each incoming video frame, a likelihood image for the object is created according to the classification results of pixels by the strong classifier. In the likelihood image the object's region turns into a blob. Different from the popular mean shift video tracking methods which determine object's size and orientation using predefined parameters, the proposed algorithm calculates objects' size and orientation from image moments of the corresponding blob, rather than trial of discrete parameters. Experiments show that the proposed algorithm achieved much better tracking precision on real video sequences than histogram based mean shift methods.