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

Abstract

In this paper, an algorithm for multiple camera based persontracking is presented. Region covariance matrixes areused to model the target appearance. The correspondencebetween multiple camera views is established via homography.It is utilized to improve the tracking of people under assumptionthat they are at the common ground plane. If thereis occlusion in one view, the homography to this view fromanother view is utilized to locate the object template. Theinformation about the true location of the template helpsthe tracker to resume, even in case of substantial temporalocclusions or large object movements. The object templateis represented by multiple non-overlapping patches. Owingto such an object representation the tracker is capable bothdetecting the occlusion and handling considerable partialocclusions. The object tracking is achieved using particleswarm optimization. The objective function is based on theLog-Euclidean Riemannian metric. Experimental resultsthat were obtained on surveillance videos show the feasibilityof the presented approach.