A GENERIC APPROACH FOR MOTION-BASED VIDEO PARSING

摘要

Motion-based video parsing methods segment video streams accordingto changes of camera motion types. They rely usuallyon 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 motioninformation is included. For such video streams, camera parametershave to be estimated using a frame-to-frame image registrationmethod. In our approach, we provide both techniques to estimatecamera parameters. Enhanced feature extraction algorithms takeadvantage of estimated parameters. For classification, the methoduses three multi-class Support Vector Machines (M-SVMs) to independentlydetect pan, tilt, and zoom camera motion as well as thedirection of motion. Experimental results show a promising performanceof our generic approach with test video streams from theTRECVid 2005 BBC rushes video corpus.