Token grouping based on 3D motion and feature selection in object tracking

A grouping method based on 3D motion and feature selection is proposed. The method uses a token with the most useful dissimilarities for grouping, selected using epipolar constraints calculated from 3D motion and a discriminant criterion. A group is extracted based on result of discriminant analysis for the selected token´s dissimilarities. The same procedure is applied recursively to remaining tokens to extract other groups. This grouping is robust because tokens with no useful information are rejected automatically. Since no nonlinear optimization is used, numerical computation is stable. In addition, no prior knowledge is needed on the number of objects. Experimental results are shown for synthetic data and real stereo image sequences