Feature Tracking by Multi-frame Relaxation

This paper describes a novel feature tracking method. It is based on an interframe relaxation technique. This method combines intra-frame and inter-frame constraints on the behaviour of acceptable contour structure. The intra-frame information is represented by both a dictionary of local contour structure and a statistical model of the response of a set of directional feature detection operators. The inter-frame ingredient represents the novel modelling component; it is encapsulated by an implicit model of the underlying surface structure of 3D feature points. The model is represented in terms of a series of unimodal probability densities whose single parameter is the interframe distance. The initial probabilities in our relaxation scheme effectively combine distributions describing the statistical uncertainties in the position and feature characteristics of multiframe contours; these probabilities are refined in the light of the dictionary to produce consistent contours. We present an experimental evaluation of the resulting feature detection method on cranial MRI data. Here the method significantly outperforms its single frame counterpart in term of its ability to extract noise-free and smooth feature contours.