Video object trajectory analysis

This article proposes a new method which allows the detection of variations of orientation of video object due to rotations around axis parallel to the image plane. For that purpose, the video object sequence is decomposed into temporal segments separated by key instants. The motion between these key instants are therefore estimated using a planar motion model. The validity of this model is tested by comparing their texture and shape. It is assumed that the considered video object is rigid and has been previously segmented for each key instant with a high quality level. If a variation of orientation has been detected, the direction of the rotation axis is estimated. Experimental results shows that this technique can classify different key views according to modification of camera/object orientation, and have an accurate estimation of the rotation axis with large motion.