184
(1) Maximum likelihood classification
The spectral feature is used for classification. Six classes are
selected: high way, car, vegetation, concrete ground, park area,
water. Figure 1 shows the discrete point graph of these six
classes. Figure 2 shows the spectral feature and figure 3 shows
a classification sample.
Figure 1. Discrete point graph in 4 dimensions
Figure 3, Image classification
(2) Vehicle detection
After image classification, the vehicle detection is performed on
the classified image based on following criteria.
■ class=2
■ Surrounded by 1
■ Not longer than 5 pixels on multispectral image
■ Mean DN number greater than a threshold
■ Mean DN number greater than its background by a threshold
Mean_DN_value=(Min+Max)/2 (1)
Grey value
The vehicle candidate’s position on the MS image is calculated
with equation 2 and 3.
II
M
(2)
n
II
(3)
Where gray(7, J) > meanDNvalue.
(3) Image matching
Because some vehicles are very close and similar (Fig. 5), small
area based matching can not differentiate these vehicles. A two
step image matching method is used to search corresponding
point on the PAN image. The first step, a large area-based
image matching is used to find a coarse position on the PAN
image. The second step takes this position as initial position,
use small area-based image matching to search the final
position.
(4) Calculation of vehicle position on the PAN image
The two-step image matching only can find the vehicle on the
PAN image. The vehicle’s central position should be detected
before calculation of vehicle position and velocity. The method
in step (2) is used to detect the vehicle’s central position on the
PAN image.
(5) Calculation of vehicle’s moving speed and direction
After the vehicle’s position on the PAN image is detected, the
vehicle’s position, moving speed, and moving direction are
calculated by a direct location algorithm (Xiong Z. and Zhang
Y„ 2007).
3. EXPERIMENT
A pair of level 1A Quickbird images which contains one 0.61
meter resolution panchromatic image and one 2.44 meter
resolution multi-spectral image (figure 3) is used to test this
technique. These images are acquired on July 26, 2002 in
Gagetown, New Brunswick, Canada. We just used a part of
whole scene image for our experiment. Below is the detailed
data clipping information.
