Digital Signal Processing Reference
In-Depth Information
Figure 4.11.
Principal Component Surface Analysis: a) by taking the first principal
component of the motion paths of the video sequence with the initial point positions,
we can derive a functional surface that represents the general movement of objects
through the video sequences. b)We can derive approximate object positions and size
through peak analysis of this surface: the bulk of coastguard ship and part of its
mast. The results were taken from frames 90-99 of the coastguard sequence.
2.
Multiple Object Separation
Our formulation of the previous section applies only to a single iso-
lated video object. For scenes with multiple objects, we must first
decompose the scene into isolated objects and apply our formulation
independently to each isolated object.
3.
Initial Shape Estimates
The containment, locality, and Voronoi Ordered Space components
of
E
object
depend upon an initial shape estimate.
To address these three issues, we apply Principal Component Analysis
to our multi-resolution optical flow to locate a conservative surface es-
timate. Our bootstrap stage finds regions of homogeneous motion that
are used as the initial conservative surface estimate,
S
initial
(see Fig-
ure 4.10).
PRINCIPAL COMPONENT MOTION FIELD
SURFACE ANALYSIS
Given a motion field from our multi-resolution optical flow, we group
pixels by their projected motion through time, treating the video se-
quence as a composition of
motion paths
(see Section 2.4). We can clus-
ter these paths by their locality and projected motion through Principal
Component Analysis (PCA) as shown in Figure 4.11. These clusters of
Search WWH ::
Custom Search