Project:
Sports Video Analysis
Current Results
Play-Break
Segmentation (For the current results, see
the project page on DVMM)
According to soccer syntax and
production rules mentioned in last section, play-break
segmentation is done by mapping grass area to view, and view sequence to game
status.
Three steps of the algorithm:
1. Grass value learning:
- Automatically tuned to the grass value
of the current clip
- Periodically updated to suit gradual change
of the field condition
- Consistent and accurate result
2. View classification
- Classifying grass area with automatically
adjusted thresholds
- Best accuracy: 92%
- Most of the errors due to model breakdown
3. Play-break segmentation
- View label sequence processing using heuristic
rules
- Different kinds of evaluation metric for
different application
- Best global accuracy: 86.5%
- Boundary timing to be improved
This result can be used for
further inference, such as grass angle classification (best accuracy of
88.5%
achieved), or combined with complementary features such as motion or field
audio to improve segmentation or do event detection.
[Reference]
P.
Xu, L. Xie, S.-F. Chang, A. Divakaran, A. Vetro, H. Sun, "Algorithms and Systems
for Segmentation and Structure Analysis in Soccer Video", IEEE International
Conference on Multimedia and Expo, Tokyo, Japan, Aug. 22-25, 2001
(PDF
)
Last
update:
06/15/2001