Covariance Region Descriptor And Energy Modeling Under Level Set Framework

Object tracking is one of the most important research topics in computer vision, which is widely used in video surveillance, medical image processing, intelligent transportation systems and other fields. In the past few decades, researchers have proposed some successful tracking algorithm, including the active contour tracking, particle filter, kernel tracking and so on. Using covariance matrix as region descriptor, in this paper we mainly study on covariance matching based active contour model and corresponding gradient descent flow.1) Using Frobenius norm of matrix to define distance metric between two covariance matrice, novel image energy is proposed that aims at minimzing covariance distance between template and the candidate object region, at the same time maximizing the covariance distance between template and the candidate background region. Two weights are added to put control over the influence of the foreground and background information on the image energy model respectively.2) By maximizing the covariance distance between the candidate object region and the candidate background region, we are still able to perform object extraction when there is lack of knowledge about template covariance beforehand, and this leads to our second image energy. The model can be also applied to image segmentation due to removing the template information.3) Motivated by Mean-Shift, a novel kernel covariance tracking is proposed. This method uses the kernel function to control the influence of individual pixel on the resulting covariance, computing object centriod in each frame by minimizing the proposed objective function.Compared with distribution matching snake, the proposed covariance matching based approach can achieve better tracking results due to its non-Euclidean geometry modeling. The adopted object representation considers various features as well as their correlation, and is not dependent on the foreground distribution and background distribution. All these make the tracking method more general. Compared with Mean-Shift, our kernel covariance method is more invariant to the mean changes such as uniform illumination and noise shift, which makes the process of tracking more stable. Experimental results show the effectiveness of the algorithm.