Extracting robust distribution using adaptive Gaussian Mixture Model and online feature selection

This paper presents a novel method to extract robust distribution using adaptive Gaussian Mixture Model (GMM) and online feature selection to improve the tracking performance. Traditional histogram-based tracking algorithms use all histogram bins to track the object, and some recent work attempt selecting important bins by directly comparing the corresponding probability value between the foreground and background histogram distribution and give higher weight to those important bins. However, the selected bins may not be discriminative enough. In this paper, we use an adaptive GMM to model histogram and select good Gaussian components corresponding to the discriminative bins and the stable part of the foreground, which form the robust distribution. Given a set of seed, for each feature, the foreground and surrounding background are modeled by an adaptive GMM, respectively. Then, we use the Gaussian Component Separability measure to find good components, to extract the robust distribution of the foreground and to evaluate the separability of each feature. Finally, after ranking features based on their separability, the K most discriminative features are used to generate a weight image and the CAMSHIFT algorithm is employed to locate the object. Experiments show that the proposed method can extract the robust distribution and improve the tracking performance.