Adaptive Kernel Tracking Based On Covariance Matching

Object tracking is one of the most important research topics in computer vision, which has wide application. This paper proposes an adaptive kernel tracking algorithm based on covariance matching in complex environment and optimizes it with the conjugate gradient method.1) In this paper, the adaptive kernel tracking algorithm based on covariance matching uses the covariance matrix to describe the target region, utilizes the kernel function to control the impact of the pixels in the target region and introduces the target motion model into the kernel function. According to Logarithmic-Euclidean distance between two covariance matrixes, our method constructs the similarity measure function between the target region and the candidate target region.As a result of introducing the target motion model into the kernel function, target tracking is converted into a parameter optimization problem. For the case of affine motion, the image model is modeled as a continuous function of the?parameters of affine transformation. By minimizing the objective function, our method gets the center, scale factor and rotation angle by iteration in each frame and updates this information in order to get an adaptive tracking window.2) Essentially, the adaptive kernel tracking algorithm based on covariance matching is a gradient descent algorithm. In order to overcome slow convergence of the gradient descent method, this paper studies the commonly used methods of unconstrained optimization. According to the basic principles and application scope of those methods, this paper selects the conjugate gradient method to optimize the adaptive kernel tracking algorithm based on covariance matching. Conjugate gradient method is established in the quadratic model with secondary termination. The performance of the conjugate gradient method is between steepest descent method and Newton's method. The conjugate gradient method can not only overcome the slow convergence of the steepest descent method, but also avoid the shortcomings of Newton method, such as large amounts of computation, local convergence and so on.A large number of experiments show that, for affine moving targets in a complex environment, the adaptive kernel tracking algorithm based on covariance matching can achieve stable and accurate tracking results, which is affected by light and noise to a lesser extent. In addition, on the aspects of convergence speed and tracking results, the conjugate gradient algorithm can achieve a significant improvement for tracking.