A generic approach for motion-based video parsing

Motion-based video parsing methods segment video streams according to changes of camera motion types. They rely usually on compressed video streams, where motion vector fields are provided. Camera parameters can be derived from these motion vectors. There are a number of relevant video codecs where no motion information is included. For such video streams, camera parameters have to be estimated using a frame-to-frame image registration method. In our approach, we provide both techniques to estimate camera parameters. Enhanced feature extraction algorithms take advantage of estimated parameters. For classification, the method uses three multi-class Support Vector Machines (M-SVMs) to independently detect pan, tilt, and zoom camera motion as well as the direction of motion. Experimental results show a promising performance of our generic approach with test video streams from the TRECVid 2005 BBC rushes video corpus.