Multi Camera-Based Person Tracking Using Region Covariance and Homography Constraint

In this paper, an algorithm for multiple camera based person tracking is presented. Region covariance matrixes are used to model the target appearance. The correspondence between multiple camera views is established via homography. It is utilized to improve the tracking of people under assumption that they are at the common ground plane. If there is occlusion in one view, the homography to this view from another view is utilized to locate the object template. The information about the true location of the template helps the tracker to resume, even in case of substantial temporal occlusions or large object movements. The object template is represented by multiple non-overlapping patches. Owing to such an object representation the tracker is capable both detecting the occlusion and handling considerable partial occlusions. The object tracking is achieved using particle swarm optimization. The objective function is based on the Log-Euclidean Riemannian metric. Experimental results that were obtained on surveillance videos show the feasibility of the presented approach.