Wide-Range Feature Point Tracking with Corresponding Point Search and Accurate Feature Point Tracking with Mean-Shift
Fumiharu Tomiyasu, Takatsugu Hirayama, Kenji Mase · 2013
We propose a Mean-Shift based feature point tracking method that can track feature points with high accuracy even when they move over a long distance or a wide range on an image. Our method selects an initial value of Mean-Shift from a wide area by a corresponding point search based on the Kalman filter when the image appearance significantly changes. The corresponding point search responds to the rapid change of the feature points because it corresponds with those detected from sequential images. We used a movement prediction of the tracking point by Kalman filter to reduce the correspondence failure. Mean-Shift search tracks the feature points accurately in a narrow range using the corresponding point as initial value. We evaluated our method by tracking feature points of synthetic image sequences that simulate a movement of the tracking target on the image. The proposed method showed smaller tracking error than both Mean-Shift search and a conventional corresponding feature point search.