Motion determination in space-time images

A new approach to determine motion from multiple images of a sequence is presented. Motion is regarded as orientation in a three-dimensional space with one time and two space coordinates. Fast cascaded filter operations analyze the orientation. The algorithm is analogous to an eigenvalue analysis of the inertia tensor in three dimensions. Four types of orientation are classified: a) constant regions, where no velocity determination is possible; b) edges, where the velocity component perpendicular to the edge is determined; c) corners, where both components of the velocity vector are calculated; d) motion discontinuities, which are used to mark the boundaries between objects moving with different velocities. The accuracy of the new algorithm has been tested with artifi cially generated image sequences with known velocity vector fields. The tests prove that accurate velocity estimates are gained. Even for a low signal to noise ratios, the standard deviation of the displace ment vectors between two consecutive images is only about 0.1 pixel distances.