Gaussian Mixture-based Mean Shift for Tracking under Abrupt Illumination Changes

Mean shift is a fundamental algorithm for visual object tracking which is based on the minimization of the distance between the discrete histogram of the target and the discrete histogram of the neighborhood of a candidate image location. While the algorithm performs well when the target's appearance and the lighting conditions are constant, it may fail when these conditions are not met because the ideal histogram is generally shifted with respect to the reference histogram. In this work, we propose to compute the initial histogram of the target using a Gaussian mixture model (GMM) rather than impulses generated by simple counting. This mixture plays the role of a weighting function, in the histograms computed in subsequent frames, in order to make them smoother and increase the overlapping area with the initial histogram. By these means, sudden illumination changes between consecutive frames may exhibit smoother transitions between the two histograms and the involved distance is not trapped into local minima.