DERIVATION OF QUALITATIVE INFORMATION IN MOTION ANALYSIS

A qualitative approach to get information about 3D kinematic behaviour of objects in a scene from apparent motion in the image sequence is presented. The process is in fact two-fold: first partition of the image into areas comprising a unique motion; and second, symbolic description of the motion of each area. Here the second part is discussed. Kinematic description involves motion and trajectory type. It relies on geometrical cues tied to the velocity field such as divergence or rotational terms. First, it is shown how a complete set of such cues can be derived through a first order development of the 2D velocity field. Second, the relation between these cues and the 3D motion parameters is established, which allows the determination of a set of labels associated with different kinematic configurations. Third, the label validation step is solved using a statistical approach (in fact, two methods have been studied). This approach avoids determining explicit 3D parameters such as a depth map or 3D motion measurements. Several experiments on different sequences are reported.