Indexing, Object Segmentation, and Event Detection in News and Sports Videos - Multimedia Database Retrieval: Technology and Applications - page 167

Section 7.6 presents a method for detecting events, and its application for the
classification of American Football videos. The starting point of play events is
detected by the MPEG-7 motion activity descriptors and the mean of the motion
vector magnitudes. Then, the descriptors from multi modalities, including MPEG-7
motion, audio descriptors, and Mel Frequency Cepstrum Coefficients descriptors,
are utilized to classify video shots into events, such as pass plays, run plays, field
goal/extra point plays and kickoff/punt plays.

7.2 Video Parsing in Compressed Domain

Video parsing is the segmentation of video sequences into video shots, which then
allows subsequent operations such as feature extraction, shot characterization, and
key frame selection. Therefore, it is crucial that video parsing can correctly detect
the shot transitions present, or else the quality of shot representation can be affected.
Since the original video data are large in size, they are compressed to preserve
storage space. The computational cost of decompression is extremely high for
most compression algorithms, and hence efficiency can be improved by performing
operations on video data in the compressed domain. MPEG is one of the widely
adopted compression standards [ 183 ], and many content-based video operations
in the compressed domain have been developed [ 180 ]. In particular, research has
shown that the DC coefficients can be used to detect shot transitions in MPDG
[ 184 ]. Previous work using Motion-JPEG indicated that AC coefficients could be
used to detect scene changes [ 185 ]. This section presents a new method called Twin-
Window Amplification Method (TWAM) [ 181 , 182 ] for detecting scene changes.
This algorithm greatly enhanced the performance of the conventional method used
in [ 184 ], and is comparable to the more complex methods in [ 186 ].

7.2.1 Conventional Method

Scene change detection is the process of dividing a video stream into shots based
on the content of the video, where a shot is a sequence of continuous frames
representing a continuous action in time and in space. In general, there are two
types of transitions: sharp transitions and gradual transitions. A sharp transition is
an abrupt transition between two shots that exists only between two frames, whereas
a gradual transition requires several frames to complete the changeover. The most
frequently encountered type of gradual transition is the dissolve transition, and it
involves the fading-out of the leading shot and the fading-in of the trailing shot.