Gaussian Mixture Model on Tensor Field for Visual Tracking
Xueliang Zhan, Bo Ma · IEEE Signal Processing Letters · 2012
Visual tracking remains a challenging problem because of both intrinsic appearance variability of object and extrinsic disturbance. To deal with this problem, we present a novel approach for tracking based on the tensor features. We convert the image into tensor field to yield more discriminating features and encode the target appearance probabilistically with gaussian mixture model (GMM). The model parameters are obtained by a modified EM algorithm using all tensor samples extracted from the target area. An incremental learning procedure is employed to update the model parameters for adapting to the appearance changes over time. Experimental results compared with three state-of-the-art methods demonstrate the good performance of the proposed algorithm under challenging conditions.