Apply GPCA to Motion Segmentation - Intelligent Video Event Analysis and Understanding - page 25

determine the occluding edges. Our development is based on the following obser-
vations (1) the intensity edges include the boundaries of motion layers; (2) the
layer boundaries are not always the occluding edges; (3) determining the occlud-
ing edges and inferring the occlusion relationship can be fulfilled by testing the
neighborhood of edges.
We will introduce the intensity edges of images into the potential occlusion ar-
eas for the detection of occluding edges in the next section.
3 Post-Processing Procedure
Let us consider a single viewpoint. The central problem is to detect the occluding
edges, because the erroneous edge labeling can cause incorrect depth ordering.
Most of the techniques considered so far employed only the motion field informa-
tion for motion segmentation. For each frame, all edges, including edges of mo-
tion layers and textured edges of objects, are presented in the image intensity
structure, which can provide the wealth of additional information to motion esti-
mation. Due to their extreme length, a number of measurements might be taken
along (or around) them. This leads to a more accurate estimation of motion.
Recent applications have motivated a renewal of motion segmentation by track-
ing edges [6,17]. Ogale et al. [7] classified the occlusions into three classes. In
order to deduce the ordinal depth, they had to fill the occluded regions. This is to
implicitly approximate the occluding edges by filling the neighborhood of the
layer boundaries. [6] provides three fundamental assumptions of the relationship
between regions and edges to identify the edges of moving regions. We add an
extra assumption (i.e. the 4 th below) and emphasize these four assumptions as
follows.
(1) As an object moves all of the edges associated with that object move, with a
motion, which may be approximately described by some motion model.
(2) The motions are layered, i.e. one motion takes place completely in front of
another, and the layers are strictly ordered. Typically the layer farthest from
the camera is referred to as the background, with nearer foreground layers in
front of this.
(3) An arbitrary segmented image region only belongs to one motion model, and
hence any occluding boundary is visible as a region edge in the image.
(4) For each frame, the intensity edges involve the edges of motion layers.
An important conclusion from these four assumptions is that the layer ordering
can be uniquely determined if the layer of each moving region is known and the
occluding edges are known. [7] presented the relationship of motion layers and
occluded regions, and further emphasized that the motion layer involving the oc-
cluded region must be behind another one. Even when the layers of motion
regions are known, ambiguities may still be presented in the layer boundary label-
ing, as shown in Fig.1. In Fig.1a, due to the occluded region C , we can infer the
occlusion relationship between the motion regions A and B , while, in Fig.1b, we