Combining discriminative and descriptive models for tracking

Abstract

In this paper, visual tracking is treated as an object/background classification problem. Multi-scale image patches are sampled to represent object and local background. A pair of binary and one-class support vector classifiers (SVC) are trained in every scale to model the object and background discriminatively and descriptively. Then a cascade structure is designed to combine SVCs in all scales. Incremental and decremental learning schemes for updating SVCs are used to adapt the environment variation, as well as to keep away from the classic problem of model drift. Two criteria are originally proposed to quantitatively evaluate the performance of tracking algorithms against model drift. Experimental results show superior accuracy and stability of our method to several state-of-the-art approaches. © Springer-Verlag 2010.