Human Tracking Using Spatialized Multi-level Histogram and Mean Shift

Sequential object tracking using mean shift method has become a convenient approach. In this method, an object of interest is represented by its global feature such as a color histogram. The next position of the target is then estimated through a constraint histogram matching. The linearization of the histogram matching metric might not work properly, especially when the target undergoes occlusion, there is an abrupt motion, or when multiple objects exist with similar global but different local structures. We propose a multi-level global-to-local histogramming approach in which the associated spatial information is also encoded in the object's representation. Specifically, for human shape/appearance encoding, the global histogram resembles the main root and the local histograms correspond to the body parts. In an experiment on a publically available CAVIAR dataset, the proposed representation provides an appropriate sequential matching of a human with abrupt motion and partial occlusion. In addition to a better localization, the proposed approach handles the situations in which the standard mean shift fails.