Affine Object Tracking Using Kernel-Based Region Covariance Descriptors

Visual tracking remains a challenging problem because of intrinsic appearance variability of object and extrinsic disturbance. Many algorithms have been recently proposed to capture the varying appearance of targets. Most existing tracking methods, however, fail to estimate the scale and orientation of the target. To deal with this problem, we model the second-order statistics of image regions using a kernel function and perform covariance matching under the Log-Euclidean Riemannian metric. Applying kernel-based covariance matrix as image region descriptor, we construct a region similarity measure that describes the relationship between the candidate object region and a given appearance template. After that, tracking is implemented by minimizing this similarity measure, in which gradient descent method is utilized to iteratively search the best matched object region. The corresponding optimization problem can be derived by calculating the first derivative of the similarity measure with respect to the affine transformation parameters and setting them to be zero. Experimental results compared with several methods demonstrate the robust performance of the proposed algorithm under challenging conditions.