Apply GPCA to Motion Segmentation - Intelligent Video Event Analysis and Understanding - page 29

2

i = arg min e p ( ) − cent ,

(2)

j

j = 1,...,8

where . On this basis, the points of W can be classified into

two groups, layer edge points and non-layer edge points.

() ( , , ()

ep eee p =

012

Region Merging
For the group of non-layer edge points, one can merge most of small spurious
regions in a big motion layer, i.e. merging small regions with a motion layer by
comparing their areas with their individual neighbors'. This can lead to the con-
nected layer. But it can be observed that the detected layer boundaries usually
have discontinuities with the group of layer edge points, i.e. a set of layer edge
segments. This is due to the fact that some layer edge points are incorrectly classi-
fied into the group of non-layer edge points. Based on the areas of the segmented
layer regions, it is impossible to make a correct decision of region merging when
these small regions may contain layer edge segments. This is because the layer
edge segments indicate that the both sides should respectively occupy different
motion layers and could not be merged into a single layer at anytime. These re-
gions are thus left as the undetermined regions temporarily.
On the other hand, a connected layer has a continuous boundary M L . These layer
edge segments only prune the region of the layer, but do not form new closed re-
gions within the layer. In our experiments, we simply replace some parts of M L with
the new layer edge segments according to the nearest neighbor criterion. Then, the
area comparison strategy is employed to those undetermined regions nearby the
layer boundary for region merging. A layer edge segment separates one region into
two motion layers. When merging two or more undetermined regions which share a
layer edge segment, the merging procedure should be terminated.

Depth Ordering
After region merging, one can obtain the desired boundaries of motion layers M L .
If the occluded regions belong to a motion layer, this layer must be behind another
one. Our problem can now be rephrased as HOW to assign the occluded regions to
the known motion layers.
With the occlusion region map M O , one can first determine which points of the
layer boundaries belong to the occluding edges, since the layer boundaries involve
the occluding edges. The worst case is that the points of the occluding edges are
not within M O . But both side profiles of these points should overlap with M O at
that moment. On this basis, one can determine the points of the occluding edges
by checking if they are within M O or their profiles overlap with M O . Since some
points of layer boundaries may not belong to the occluding edges, such as in
Fig.1b, the depth ordering can only carry out on the detected occluding edges.
Then, for the points of the occluding edges, one can extend their profiles in the
direction of the intensity gradient to the known motion layers for their profile
labeling, i.e. inferring which layers both sides of the occluding edges respectively
belong to.