Dr. Suchi Bhandarkar
 
Graduate Students

Aparna Kadakia
Yash Warke
 
    The emergence of multimedia information systems has created a need for analysis, integration and organization of non-traditional data such as digitized images, video and audio. One of the greatest challenges in the design of multimedia information systems is overcoming the difficulty of rapidly and reliably extracting ``key'' information from images, video and audio streams which could then be used for rapid browsing and indexing of the underlying information. In this project we focus on  the analysis of compressed (MPEG) video data with the intent of extracting suitable key information.
 
    Video parsing or scene/shot change detection in a video stream is commonly used to extract key frames in a video stream. These key frames are then used for rapid video browsing and automatic annotation and indexing of video streams to support content based query access to large video databases. The video parsing operation is primarily domain independent i.e., no assumptions are made about the semantics of the video or its underlying theme. Video parsing, therefore, is a crucial first step that precedes domain dependent analysis of the video. Due to the large amount of data involved, video streams are often compressed for efficient transmission and storage. Video parsing techniques that are capable of processing compressed video data directly have a considerable advantage in terms of execution time and memory requirement over those that require full frame decompression. Consequently, this project focuses on the design and implementation of video parsing and video analysis techniques that are capable of processing compressed data directly.

          The figure shows the motion vectors used in the algorithm
 
    The research thus far has resulted in the design and implementation of a video parsing algorithm for MPEG (compressed) video data. The algorithm integrates motion cues and chrominance/luminance cues from the video data. Motion cues are in the form of motion vectors which are derived from the motion compensation information encoded in the MPEG stream. Chrominance and luminance cues are derived from the DC images in the MPEG stream. The algorithm detects abrupt changes (cuts or shot boundaries), gradual scene changes (fades and dissolves) and camera motion parameters such as pans and zooms. Experimental results on MPEG video showed the algorithm to be fast (~60 frames per second on a 170 MHz SUN UltraSPARC1 workstation) and accurate.

          The figure shows the functionality of the algorithm

    Current work in this project deals with key frame generation and representation in video sequences in particular, generation of video mosaics from compressed video, annotation of mosaics with motion based information and extraction of indexing attributes.
 
Publications
 
S.M. Bhandarkar and A.A. Khombadia, Motion based Parsing of Compressed Video, Proc. IEEE Intl. Wkshp.  Multimedia Database Mgmt. Sys., Dayton, Ohio, August 5-7, 1998, pp. 80-87.
 
A.A. Khombadia, A Rapid MPEG Navigator, MS Thesis, Dept. of Computer Science, University of Georgia, 1997.
 
Y.S. Warke, Integrated Parsing of Compressed Video, MS Thesis, Dept. of Computer Science, University of Georgia, 1998.