Joint Gaussian Conditional Random Fields for Background Extraction [PDF]
Abstract
Background extraction is the important technique for video processing applications.
Most existing methods tend to be unsuited when solving on complex foreground movement
and surrounding illumination changes video sequences. To extract background from these
video sequences, we introduce a novel background extraction method based on joint Gaussian
conditional random fields (JGCRF) probabilistic model of fixed-view video sequences.
The model analyzes the intra-frame and inter-frame relationship to reduce the visual
artifacts coming from surrounding illumination changes. Additionally, a motion-less
patch extraction technique based on the sum of squared distance between corresponding
pixels in two consecutive frames is proposed to solve the complex foreground movement
problem. As shown in the experiments, our method achieve better background extraction
results compared to the existing methods. Further, JGCRF can achieve high accuracy and
high speed simultaneously when dealing with full high-definition format video sequences.
Results
Low resolution videos
Double-click the images to enlarge them
and click once to make them thumbnail size again.
|
Medium |
PBI [1] |
Hybird [2] |
GMM [3] |
Ours Naive |
Ours JGCRF |
ved |
|
|
|
|
|
|
man |
|
|
|
|
|
|
car |
|
|
|
|
|
|
board |
|
|
|
|
|
|
lab |
|
|
|
|
|
|
road |
|
|
|
|
|
|
fish |
|
|
|
|
|
|
High resolution videos
Double-click the images to enlarge them
and click once to make them thumbnail size again.
|
Ground Truth |
Medium |
PBI [1] |
Hybird [2] |
GMM [3] |
Ours Naive |
Ours JGCRF |
HD1 |
|
|
|
|
|
|
|
HD2 |
|
|
|
|
|
|
|
HD3 |
|
|
|
|
|
|
|
HD4 |
|
|
|
|
|
|
|
HD5 |
|
|
|
|
|
|
|
HD6 |
|
|
|
|
|
|
|
HD7 |
|
|
|
|
|
|
|
HD8 |
|
|
|
|
|
|
|
HD9 |
|
|
|
|
|
|
|
Videos
ved
man
car
board
lab
fish
HD1
HD2
HD3
HD4
HD5
HD6
HD7
HD8
HD9
References
- [1] A. Colombari, and A. Fusiello.
Patch-based background initialization in heavily cluttered video,
Image Processing, IEEE Transactions on, vol. 19, pp. 926-933, Apr. 2010.
- [2] C.-C. Chen, and J. K. Aggarwal.
An adaptive background model initialization algorithm with objects moving at different depths,
Image Processing, IEEE International Conference on, 2008.
- [3] C. Stauffer and W. Grimson.
Adaptive background mixture models for real-time tracking,
in Proc. Conf. Comput. Vis. Pattern Recognit., vol. 2, pp. 246-252, Jun. 1999.