Indexing, Object Segmentation, and Event Detection in News and Sports Videos - Multimedia Database Retrieval: Technology and Applications - page 206

Table 7.3 Play event classification results, obtained by multiple feature types

Play
category

MPEG-7 audio
+

MPEG-7 motion
+

MPEG-7 motion
+

MPEG-7 motion +

MFCC audio MFCC audio + MFCC

Pass 70.0 % 85.2 % 85.2 % 94.3 %

Run 59.7 % 91.0 % 92.5 % 89.6 %

FG/XP 75.0 % 87.5 % 87.5 % 93.8 %

K/P 69.0 % 82.8 % 82.8 % 93.1 %

Overall 67 . 0% 87 . 0% 87 . 5% 92 . 5%

Table 7.4 Play event classification results, obtained by three sets of
features, based on motion combined with other modalities
Method Pass Run EG/XP K/P

MPEG-7 motion 79.5 % 92.5 % 87.5 % 65.5 %

MPEG-7 motion + audio 85.2 % 91.0 % 87.5 % 82.8 %

MPEG7 motion + audio + MFCC 94.3 % 89.6 % 93.8 % 93.1 %

different networks. This variety in the database ensured that the sample space of the
current work was diverse and included all the major broadcasters.
Table 7.3 , shows the indexing results of using MPEG-7 motion and audio
descriptors along with MFCC features. From table, we can see the classification
accuracy increased with the combining of multi-modal features. In the case of
combining the MPEG-7 audio with MFCC features, we see an overall increase of
10 %, while combining the audio features with motion descriptor features shows
an increase of 5 %. Combining all three features produces an overall classification
result of 92.5 %.
Combining multi-modal features in a reasonable fashion can enhance the
classification. But always there are trade-offs that need to be considered. Some
features may reduce the accuracy of classification of a particular category but may
enhance the overall performance of the system. Table 7.4 shows the variations in
classification that results from adding audio features to the motion features.

7.7 Summary

The chapter covers a broad spectrum of video segmentation, indexing, retrieval,
and classification techniques applicable to news and sports videos. Based on the
energy histogram of DCT coefficients, a shot detection algorithm for MPEG video
data in the compressed domain can be developed. The detection results can be
enhanced by using the ratio between two sliding windows to attenuate the low-pass
filtered frame distances. The advantage is in achieving high detection rates with low
computational complexity. In a subsequent process, news videos can be segmented
into shot, group-of-shots, and story levels, where the template frequency model can
be applied to capture the spatio-temporal information. This facilitates video retrieval