g to provide useful
vo computer vision
ie second is optical
grated. The results
hip
riments
to detect and track
1 à set of sequential
ir target scenes. To
aches are known in
e (call it template)
ract it from each of
ected positions can
reliable extraction
n problem if the
Part 5. Hakodate 1998
situation is dynamically changing (typically appar-
ent size will change). Also, the premise that the object
shape is previously known is not realistic because to
prepare the templates for a variety of various ships is
4 difficult problem. Though the difference operator
adopted to a set of sequential images makes position
information of moving objects, reliable tracking of a
ship is still an open problem. The specific situation
for detection of a moving ship is that it will cause a
wake following to the ship image. The difference oper-
ator will detect the wake part in addition to the ship
part. However the optical flow method [4][5] is compu-
tationally expensive, the problems in above mentioned
methods are relatively easier to be solved. The ex-
tracted velocity vector for each pixel will be useful
to obtain the movement of observer, to obtain rela-
tive depth map, and to reconstruct 3D information of
the ship. In this paper we propose a scheme in which
the optical flow method is used mainly and the mod-
ified template matching method is also used. It is in-
tended to extract ship part from a set of (supposed
to be) uniform velocity vectors calculated by the op-
tical low method. The template matching method is
used to know macroscopic movement of a ship, whereas
the optical flow provide microscopic movement of each
pixel. The template used here is not previously pre-
pared, but extracted from the image sequence. It is
devised to classify the velocity vectors (detected by the
optical low method) into ones belonging to ship part
and other part by using the macroscopic movement in-
formation obtained from the template. The computer
programs we made and used are for extracting tem-
plate, for tracking of a ship, and for calculating optical
flow, and for extracting ship area. They are described
in the following sections.
3. Template matching and tracking
It is a difficult problem to prepare the templates for
a variety of various ships. The template we need here
is just for tracking macroscopic movement of the ship.
Therefore, we do not use the previously prepared tem-
plate, but use ones extracted from the image sequence.
The bow part of the ship is defined to be the tem-
plate here. The difference operator is adopted to the
n-th frame and (n+i)-th frame in the image sequence,
followed by the thresholding operator. The obtained
binary image is projected to the horizontal and verti-
cal axes (see Figure 2). We decide that the one axis out
of two axes is closer to the direction of ship movement
(call the moving direction) if the projection width is
wider. The horizontal axis is considered to be closer
to the moving direction in figure 2. The width of the
projection perpendicular to the axis mentioned above
(a in figure 2) is used as a measure of bow part of
the ship. The pixels within the rectangle defined by
a and 2a are used for the template, where the start
point of 2a is defined as a part where the cumulative
sum of projection is growing. In figure 3, shows an ex-
ample of a template. For each image in the sequence,
the template is decided as described above. The n-th
frame and (n+i)-th frame are compared and position in
n+i th frame where the template in n-th frame mostly
matches with one in n+i th frame. The following sim-
ilarity measure is used:
g(x,y) = [20 [Zoo P(x — uy — v)f(u, v)dudv
K J [$91 p^(u, v)dudo - y= jen: f?(u,v)dudv
d'(x,y)= ko(z,v)
where
f(z,y):image of a ship
p(i, t):template
g' (i, t):similarity measure
iG. s 1.
A set of detected positions for bow of the ship pro-
vides tracking information of the ship movement.
rp
2a
aad 2 Leia d aa
Figure 2: Extraction of a template
Figure 3: Extracted template
