Multi-scale object tracking based on mean shift and AUC

The mean-shift algorithm is an efficient technique for 2D object tracking. Although the scale of the mean-shift kernel is a crucial parameter, there is presently no clean mechanism for choosing or updating scale while tracking objects that are changing in size. In this paper, Lindeberg's theory of scale selection based on local maxima of differential scale-space filters is improved and then adapted to select suitable scale of the mean-shift kernel in the process of multi-scale object tracking. In addition, AUC is regarded as a standard to evaluate the efficiency of the algorithm put forward above.