F. Mokhtarian and F. Mohanna (UK)
corner detection, curvature scale space, canny edge detector, corner matching, multiple hypotheses
This paper presents a corner tracking algorithm which
maintains multiple hypotheses while tracking corners.
To extract corners from each frame of video sequences,
the multi-scale enhanced CSS1
corner detector is applied.
In the matching stage two-frame correspondence using multiple matching combined with three-frame based
monitoring is employed. Considering multiple hypotheses helps to ensure that if the corner detection is not completely robust, we still have a chance to ﬁnd as close as
possible a match or multiple matches for tracked corners. False matches are removed using a distance criterion based on a reference point. The value of threshold in
the distance criterion is constant and does not need any
adjustment. Since proposed tracker does not make any
assumptions or use any special motion models in tracking feature points, it is more practical and efﬁcient for unconstrained and non-smooth motions. Experiments have
been carried out on a wide range of real video sequences.
All experiments conﬁrm that the proposed corner tracker
performs reliably and more independently of corner detector robustness.