Motion-based video indexing evaluating object shading

This paper proposes a motion-based video sequence indexing method for scenes with non-uniform illumination, i.e. object shading and cast shadows may be present. The motion feature used for indexing is the 3D object rotation axis projected into the image plane that is defined by a tilt angle. As previous work shows, the tilt angle can be easily estimated from a video sequence evaluating only temporally changing shading on object surfaces-an effect that is often neglected in computer vision. A new indexing method is proposed that cuts the video sequence into temporal segments that are characterized by a tilt angle following the homogeneous motion model of a temporal Kalman filter. Experiments with real video sequences of simple objects like a ball or a persons head show that the video is cut into semantically meaningful segments