Apply GPCA to Motion Segmentation - Intelligent Video Event Analysis and Understanding - page 34

frame 1 frame2 1. background 2. arm region
3. ball region 4. Background 5. arm region 6. ball region
Fig. 6 The segmentation results with 3 groups. 1)-3) are the results by the subspace seg-
mentation method; 4)-6) are the results of the post-processing procedure.
However, this segmentation result is not unique in this sequence. In the other
two successive frames, the ball and arm are classified into the same group, which
is shown in Fig.7. Although the region of the ball is disconnected to that of the
arm, they are still regarded as sharing the same affine motion model. This indi-
cates that motion segmentation cannot guarantee the uniqueness of the solution.
frame 1 frame 2
1. background 2. Foreground 3. background 4. foreground
Fig. 7 The segmentation results with 2 groups. 1)-2) are the results of subspace segmenta-
tion method; 3)-4) are the results of the post-processing procedure.
5 Conclusion
In this paper, we proposed a novel approach for motion segmentation based on the
subspace segmentation techniques. The novelty is that by incorporating the intensity
structures of images, our proposed approach can effectively detect the motion layer
boundaries and the depth ordering. Different from the previous motion segmentation
approaches, our approach provides a non-iterative and global solution to motion
segmentation under a unified algebra framework, i.e. the generalized PCA [12,13].