Object Tracking Algorithms Based On Covariance Matrix

Object detecting and tracking in image sequence has been widely used in many computer vision applications, and it has extremely important theory significance and practical application value. As a kind of feature descriptor, Covariance matrix has strong ability in feature description, and has certain robustness with low dimension, because of its natural integrating with variety of features. So it has been focused on in the field of object tracking.In this paper, we based on the covariance matrix, aiming at dealing with some defects in traditional covariance tracking algorithm. Because of the manifold properties of covariance matrix(covariance matrix is one of the SPD manifold), we combine with compressed sensing theory, and study two kinds of the tracking algorithm based on covariance matrix:First, consider of the traditional algorithm in object tracking based on covariance matrix, they use to spend too much time in calculating integral image and searching object globally. So we propose a covariance tracking algorithm based on adaptive region of integral area, based on the current tracking performance of detection, dynamic adjust the search area, and calculate on the search area, thus in normally tracking, we reduce the amount of calculation, speed up the tracking; When we have tracked a shelter, by automatically expanding the search area to the whole image, we back to the traditional covariance matrix algorithm processing, guarantee the robustness of tracking. This algorithm can greatly increase the speed without losing robustness and accuracy.Second, combines object tracking algorithm based on compression sensing and covariance matrix, we have proposed covariance tracking algorithm based on compress sensing. First using compresse sensing principle for compressing Haar features of the object area, then using covariance matrix to mix the underlying multidimensional features in Haar feature area, at last using the covariance matrix tot model object, by searching the neighbor field in the object area, search the best matching to the current object model, finally using manifold space mean updating strategy to update the object model to improve the algorithm accuracy. The experimental results show that the algorithm has better tracking accuracy.