Adaboost video tracking

Abstract

A new approach of tracking objects in image sequences is proposed, in which tracking is seen as a binary classification problem. For each incoming image frame, a likelihood image for the object is created according to the classification results of pixels by a Adaboost feature classifier. In the likelihood image the object's region turns into a blob. The scale of this blob can be determined by the local maxima of differential scale-space filter. We employ the QP_TR trust region algorithm to search for the local maxima of the multi-scale normalized Laplacian filter of the likelihood image so as to locate the object as well as determine its scale. The object's appearance change is dealt with in the update step of the feature classifier. Based on the tracking results of sequence examples, the novel method has been proven to be capable of describing the object more accurately and thus achieves much better tracking precision. © 2008 Springer-Verlag Berlin Heidelberg.