Multiphase vs. single-phase variational level set approach for video data association

Abstract

In general, a multi-object tracking problem in a network of cameras consists of two steps. First, objects are tracked in each camera and tracklets, which are the traces of each object in each camera, are extracted by any common tracking algorithm. In the second step, the extracted tracklets are associated and persistent trace of the objects are obtained. In this paper, our focus is concentrated on the second step where we propose a multiphase variational method and compare it against single-phase variational model applied for the association. Both methods use the extracted tracklets as the input and the persistent trace of objects are evaluated by associating the corresponding tracklets of objects. The association is formulated by optimizing a variational energy function, which is constructed based on appearance and motion model of the objects. The optimization problem is solved by, first converting the variational energy function into an Ordinary Differential Equation (ODE) employing the Euler-Lagrange equation; then solving the obtained ODE through numerical methods. Experimental results on the real and synthetic datasets demonstrate that the multiphase model yields more complete tracking results, while the single-phase model yields more accurate tracking results.