Visual tracking based on online sparse feature learning

Abstract

Various visual tracking approaches have been proposed for robust target tracking, among which using sparse representation of the tracking target yields promising performance. Some earlier works in this line used a fixed subset of features to compress the target's appearance, which has limited modeling capacity between the target and the background, and could not accommodate their appearance change over long period of time. In this paper, we propose a visual tracking method by modeling targets with online-learned sparse features. We first extract high dimensional Haar-like features as an over-completed basis set, and then solve the feature selection problem in an efficient L1-regularized sparse-coding process. The selected low-dimensional representation best discriminates the target from its neighboring background. Next we use a naive Bayesian classifier to select the most-likely target candidate by a binary classification process. The online feature selection process happens when there are significant appearance changes identified by a thresholding strategy. In this way, our proposed method could work for long tracking tasks. At the same time, our comprehensive experimental evaluation has shown that the proposed methods achieve excellent running speed and higher accuracy over many state-of-the-art approaches.