Dynamic feature point tracking in an image sequence

This paper presents a model-based algorithm for tracking feature points over a long sequence of monocular noisy images with the ability to include new feature points detected in successive frames. The trajectory for each feature point is modeled by a simple kinematic motion model. A probabilistic data association filter is first designed to estimate the motion between two consecutive frames. A matching algorithm then identifies the corresponding point to sub-pixel accuracy and an extended Kalman filter (EKF) is employed to continuously track the feature point. An efficient way to dynamically include new feature points from successive frames into a tracking list is also addressed. Tracking results for two image sequences are given