Image-based multi-target tracking using a multi-layer particle filter and extended EM clustering

The paper presents an approach for tracking a variable number of objects by using a multi-layer particle filter combined with an extended Expectation Maximization (EM) clustering. The approach works on basis of binary foreground images coming from previous background subtraction. The multi-layer particle filter is an improvement of a conventional particle filter approach. It uses an introduced adaptive layer distribution spanned over the tracking area, which determines the areal extents of the particles. Thus, the multi-modal posterior distribution representing the objects is approximated with locally different resolutions. In addition, the layer distribution is used to find new appearing objects. In order to generate an object list out of the particle density, an EM clustering is used. The basic algorithm is extended with an estimation of the needful number of clusters by iteratively splitting and comparing the overall cluster areas. The new tracking approach improves tracking quality and robustness compared to the conventional particle filter approach. Experimental results are shown using the example of a traffic scene in a roundabout.