Object compressive tracking via online feature selection

The compressive tracking algorithm based on compressive sensing theory can efficiently achieve real-time object tracking, but the algorithm does not select proper object features online, resulting in low tracking robustness. In order to solve this problem, an object compressive tracking algorithm with online feature selection is presented. Firstly, sets of positive and negative samples are obtained by sampling around the object, and the multi-scale rectangle features of the samples are calculated. Secondly, the compressive sensing random projection matrix is used to reduce the dimensionality of high dimensional features to obtain low-dimensional compressive domain features, and the compressive domain features are updated and selected online to extract the optimal feature to remove contaminated samples and update the classifier. Finally, a simple and efficient Bayesian classification model is utilized to achieve the object tracking. Moreover, changes of object scale in the camera are modeled and a quantitative description of changes in scale is given for multi-scale tracking which can adapt to change of the object scale. Experimental results show that the proposed algorithm can achieve a higher tracking accuracy and better robustness than several state-of-the-art algorithms and can well respond to the interferences such as block in the object tracking, light mutation, scale changes, non-rigid deformation and so on. Meanwhile, it has a low computational complexity and fully satisfies the real-time requirement.