Mean-Shift Tracking for Surveillance: Evaluations and Enhancements

Abstract

Mean-shift tracking is a technique for following an object of interest as it moves through a video sequence. It is a gradient ascent approach that models the image region to be tracked by its colour histogram. In this thesis, we apply mean shift in the domain of surveillance in order to track people as they walk through a scene. Our objectives are to evaluate the performance of the technique and subsequently to introduce modifications which make the method more robust, i.e., more likely to follow a designated target through an entire video sequence.

We first compare mean shift to a standard template matching approach. The latter is found to be much more reliable, rarely losing track of its target, and so its performance serves as a baseline against which to measure the effects of our subsequent modifications to the basic mean-shift method.

In an effort to improve the reliability of mean shift, we employ an existing technique ? the use of multiple-part models ? to introduce a degree of spatial structure into its histograms, mimicking one of the strengths of template matching. We further extend the method by exploiting background models of the scene, another widely used modification. Our innovation of combining the two enhancements, while not enabling mean shift to reach the performance of the template matching tracker, increases its reliability considerably.

There are several parameters associated with any tracker. In the case of using the meanshift technique in the surveillance domain, we seek the optimal choices for the colour space in which it operates and the size of its model histograms, among other parameters. Once again, the evaluations allow us to improve the performance of the method.

The greatest advantage of mean shift over other tracking techniques is arguably its computational efficiency (deriving from its gradient ascent nature), even if this comes at the expense of lower robustness. However, we succeed in developing a tracker based on normalised cross-correlation (the similarity measure at the heart of template matching) that also uses a gradient ascent optimisation strategy. The new approach is both fast and significantly more reliable than mean shift, calling into question the use of the latter technique.

We also define an algorithm for verifying that the output of a tracker is trustworthy. Unlike our previous evaluations, the track validation approach does not require the use of ground truth data, and provides a semi-automated means of assessing the performance of any tracking method. It also allows our new gradient-based tracker to update its model in response to appearance changes in the target, further increasing the reliability of the method.