MIL based visual object tracking with kernel and scale adaptation

Abstract

Multiple instance learning (MIL) is a framework wherein training examples are provided in form of labeled bags rather than labeled instances. For visual object tracking, the MIL framework is used to select a few discriminative features from a pool of features by maximizing the bag likelihood. The feature pool consists of computationally efficient Haar-like features. Each feature is used for the construction of weak classifier also called decision stump. A few classifiers with high discriminative power (to separate target from background) are combined to form a strong classifier. The key point is the selection of a few highly discriminative features from a very large pool in small amount of time to ensure fast and accurate tracking. In this paper, we propose a fast online feature selection algorithm based on maximizing the classifier score (CSR). The tracking performance is better than the existing feature selection methods. This paper also introduces kernel trick on Haar-like features for target tracking. Furthermore, we explore use of Haar-features in half target space. To get the actual target size in successive video frames, a scaling strategy is applied. By having the feature matching using kernel, Haar-features in half target spaces and scale adaptation method in a single tracking framework, we obtain better tracking performance as compared to the state-of-the-art trackers.