Feature correspondence using probabilistic data association

A complete algorithm for feature point correspondence of a long sequence of images is presented. First, feature points are extracted from the first frame. Then, based on a 2-D constant translation and rotation model, an extended Kalman filter is used to predict the location of the corresponding point. Matching is done by comparing the feature vector and a motion continuity measure. Track initiation and termination are handled by the probabilistic data association filter. A method for including new features before the termination of gradually unreliable trajectories is introduced. Experimental results are presented for two real image sequences: a NASA helicopter sequence and a PUMA sequence.<>