Digital Signal Processing Reference
In-Depth Information
Fig. 5.10
Depth-assisted object segmentation: (
a
) initial frame, (
b
) initial labeling by depth clus-
tering, (
c
) improved classification using depth ordering, (
d
) objects segmentation results,
top
:
Three-People sequence,
bottom
: IU sequence
in Fig.
5.10
c, graph cut with
-expansion is employed to minimize the energy
function by fusing the color and motion cues, and segment the multiple human
objects simultaneously.
α
5.3
Multiview Video Tracking
In the video sequence, observers generally focus more on the static or moving ob-
ject than the background, and the video-based applications are commonly developed
based on the object-level, thus we mainly address the object-based multiview video
tracking. Comparing with the image segmentation, segmenting the sequence into
spatiotemporally consistent volume and tracking meaningful video object across
frames increase the complexity of video segmentation and tracking. Segmenting
and tracking multiple moving objects automatically and accurately in wide-range
environment or with largely crowded individuals is more challenging task. Using
monocular camera is insufficient due to the limited field of views and significant
occlusion. stereo/multiple cameras are reasonable alternatives to solve this difficult
problem by data association across-time and across-view. The current works on
multiview video segmentation and tracking are categorized into two main classes,
namely feature-based and homogrophy-based, where segmentation of video objects
are described as different representations such as bounding box, 2D/3D ellipse, or
foreground mask. The categorical overview of multiview video segmentation and
tracking approaches is shown in Fig.
5.11
. Feature-based approaches perform seg-
mentation and tracking by fusing various 2D/3D features and/or introducing statistic
filters. Homogrophy-based approaches incorporate homogrophy constraint across
multiple views into the segmentation and tracking framework.
