Color Information Fusion For Visual Objects Tracking

An efficient and accurate method for selecting features is important in tracking object. In this paper, an online feature selection approach is proposed to improve the discrimination between objects and background which makes the objects distinct from the background in the feature distribution. In this paper, by fusing the multi-channel images, the object information is illustrated in a multiple dimension feature space. We use Fisher linear discriminant analysis to reduce dimensionality for finding the most discriminative projection vector, on which the projecting result shows low intra-class variance and high inter-class variance with a higher discrimination. This scheme is embedded into mean shift tracking system and the experimental results show our scheme can enhance the robustness, real-time performance and the accuracy.