Recursive tracking of image points using labelled graph matching

The authors address the problem of obtaining the image-plane trajectories of feature points from a sequence of images. The matching of points between successive images in the sequence is treated as a labeled graph matching problem, and solved by minimizing a weighted sum of three cost functions. The nodes in the graph are the feature points to be matched, which are labeled using Gabor wavelets. Information gathered from the previous image frames is processed by a Kalman filter and used to predict the positions of the match points in the incoming image, and thereby initialize the graph matching. Results on synthetic and real image sequences are presented