Tracking based solely on location of interest points

This paper proposes two algorithms for tracking a region of interest. Given a fairly small number of spatial points at each frame and a high level of outliers, a bounding box is drawn around the region of interest. The first algorithm searches the bounding box having the greatest density. The second algorithm searches the bounding box whose theoretical 2D random distribution matches best with the 2D empirical random distribution. An estimate of the error is also provided by the second algorithm. These two algorithms are tested on synthetic data and real data: the second algorithm has better performance on synthetic data, whereas the first algorithm has better performance on real data and could be used in real-time applications. As information related to a feature can be represented as a set of points, these algorithms may prove useful to computer vision designers in feature selection or in data fusion (fusion of different tracks derived from different features).