el
Ss
d
to
ly
y
al
1€
Fuse, Takashi
4.2.2 Application to a Sample Image
We also applied the above-mentioned four methods to a sample image produced from an aerial HDTV image (Figure 3).
The data of this image are as follows:
Platform: Helicopter;
Time: 1996, March 16th;
Area: Yodogawa-River Area;
Length of the road: about 700m (one-way and two-lane);
Altitude of platform: about 500m;
Time interval of successive image: 1.5 seconds;
Spatial resolution: 0.33m ;
Number of channels: 3 (R, G and B, 8bit);
Size of image: 1920 by 650 pixels;
Number of detected vehicles: 47 in Image 1,
48 in Image 2 (one vehicle appeared).
Figure 3: Sample Image.
We used the value of parameters specified in the Section 4.2.1. We applied the four methods to both of types of
images as mentioned in Section 4.1.
Vehicle detection was accomplished manually. The coordinates of vehicles were the center of gravity, and the colors
were the average of the each vehicle. For calculating the correct rate, accurate labeling were constructed with
successive images at a short interval manually.
Table 2 shows the results with the sample image. The proposed method produced improved results than the original
method, and the rate of correct correspondence is above 9596. In the case of (c) with opposite label, the results were
better than (a) original method and (b) with color information. However, they were much worse than those by the
proposed method, because there existed vehicles whose colors were not similar and displacement vectors were similar.
The algorithm of relaxation method improves the probabilities using only consistency property. So, if these vehicles
exist, the label probabilities cannot reach convergence. The proposed method has advantage over (c) with opposite
label method due to use color information.
Table 2: Correct Rates to Sample Image
(1) Without | (2) With
(a) Original method 78.7% 75.6%
(b) With color information 80.9% 75.6%
(c) With opposite label 83.296 78.996
(d) Proposed method 95.8% 96.7%
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B5. Amsterdam 2000. 283