Leveraging observation uncertainty for robust visual tracking

Our main contribution is to enhance trackers by leveraging observation uncertainty in terms of mutual information.As another contribution, we theoretically underpin that the observation uncertainty problem can be resolved by three image processing techniques: background subtraction, saliency detection, and extracting illumination-invariant (intrinsic) intensities.Our approach can be plugged into any kind of visual trackers. In this paper, the accuracy of visual tracking is enhanced by leveraging a novel measure for observation quality. We measure observation quality with mutual information, then look at the interval covered by that mutual information. As observation uncertainty the interval length is proposed. The best observation is considered the one that both maximizes the observation quality and minimizes the observation uncertainty. We show that searching for the best observation in these terms amounts to preprocessing the image by subtracting the background, detecting salient regions, and rendering the image illumination-invariant. These preprocessing steps are very fast and can precede any existing tracker. In experiments it is shown that the performance of several trackers can be substantially boosted when they run on our preprocessed images, rather than the raw input for which they were intended. In all cases the version with preprocessing significantly outperforms the original trackers performance.