Digital Signal Processing Reference
In-Depth Information
If a foreground region are partially overlap a vehicle previously detected, assign
the corresponding old vehicle ID to the blocks in the foreground region. Otherwise,
give them a new vehicle ID.
3.2
Tracking Vehicles to Next Frame
The motion vector of a block is the result of the coordinate of the matching block in
the next frame image subtracting the coordinate of the current block. The matching
extent between two blocks is examined by the
MAD
(mean absolute difference) be-
tween them in this paper. The
MAD
is calculated as Formulas (1).
1
MAD B
[
(
x
,
y
),
B
(
x
,
y
)]=
I
( ,
x y
) -
I
(
x
+
dx y
,
+
dy
)
(1)
k
-1
1
1
k
2
2
k-1
k
64
(,)
xy
∈
B
( , )
x y
k
-1
1
1
where
B
k-1
(x
1
,y
1
)
represents the block whose coordinate is
(x
1
,y
1
)
in the k
th
frame im-
age,
B
k
(x
2
,y
2
)
represents the block whose coordinate is
(x
2
,y
2
)
in the k
th
frame image,
MAD[B
k-1
(x
1
,y
1
), B
k
(x
2
,y
2
)]
represents the
MAD
between
B
k-1
(x
1
,y
1
)
and
B
k-1
(x
1
,y
1
)
,
I
k-
1
(x,y)
represents the intensity of the pixel at
(x,y)
,
(dx,dy)
represents
(x
2
-x
1
,y
2
-y
1
)
.
According to the principle of the least
MAD
, the motion vector of a block is gotten
as Formulas (2).
(
VV
,
)
=
arg min
MADB x y Bx
[
(
,
),
(
+
dxy
,
+
dy
)]
(2)
x
y
11
1
1
1,1, -1
k
k
-1
k
(,
dx dy
)
where
(V
x
,V
y
)
1
,
1
,
k-1
indicates the motion vector of
B
k-1
(x
1
,y
1
)
.
Shift all vehicle blocks to the next image based on the pixel level motion vector es-
timated by block matching, and assign them the corresponding vehicle ID. Unless
different blocks in the k
th
frame image enjoy the same matching block in the k+1
th
frame image, all the blocks belong to vehicles can be inherited in next frame image.
3.3
Correcting Errors of Motion Vectors
Although pixel level motion vector is more accurate than block level motion vector,
the block matching method can't assure all motion vectors are right. Sometimes, some
errors occur in bounding blocks of vehicles, where very poor texture exists. Since
those errors should lead to segmentation errors, it is necessary to correct errors of
motion vectors after shifting vehicles to the next frame.
At first, check whether each block with a vehicle ID is a background block or not,
give the background blocks the background ID. Thus, a lot of wrong matching blocks
can be filtered out as shown in Fig. 4. However, if a block of one vehicle in the k
th
frame image is wrongly matched with a block of another vehicle in the k+1th frame
image as shown in Fig. 5(a). The background block verification method for correct
errors of motion vectors described above will lose efficacy. In fact, the segmentation
errors can be easily corrected just by giving the overlapping foreground blocks among
different vehicles the background ID as shown in as shown in Fig. 5(b). Thanks to the
high blocks inheriting rate achieved in the process of shifting vehicles, the wrongly
matched blocks are abandoned vehicle IDs safely, and using the remained vehicle
blocks still can keep vehicles tracked.
