Ree. — 4 5 TY, 
m fd 
T 
f 
€ UU 
» 
OR FA Oo 
Fuse, Takashi 
  
as follows in applying to vehicle tracking (Barnard and Thompson, 1980, Takagi and Shimoda, 1991). We used two 
successive images, Image 1 and Image 2. At first, we select candidates in Image 2 correspond to a vehicle in Image 1. 
The candidates are restricted by the velocity of the vehicle. These are referred to as possible correspondences. They 
include no correspondence due to the disappearance of the vehicle. And then, probabilities to the correspondences are 
set equally. The probabilities are improved successively by applying a consistency property, that is the similarity 
between displacement vectors of near vehicles. Finally, the most possible correspondence will have the highest 
probability. 
However, the probabilistic relaxation method cannot take the color of vehicles into account, which is very significant 
information, and it cannot be applied to the appearance of vehicles. Taking countermeasure against these problems, 
we improve the probabilistic relaxation method by introducing: 
(a) the color information of vehicles , and 
(b) the displacement vectors of each other. 
(1) Specifying Initial Probabilities 
Color information is very significant for correspondence between vehicles. So the color information is introduced to 
specifying initial probabilities of the possible correspondences, namely the initial probabilities of correspondences are 
varied according to similarity of the colors. 
From now, we introduce abbreviations as follows: 
r,(x, y), g,(<, y), b,(«, y): Intensity of R, G, B (8bit) in Image 1, 
r(x, y), gx, y), b.(x, y): Intensity of R, G, B (8bit) in Image 2. 
In both images, the value of intensity is 
  
0<rg,h=255= Char ‘ (1) 
D: Maximum of vehicles movement (pixels). 
a; Vehicles detected in Image 1 locate (x; y). i=1, 2v , M. 
b,: Vehicles detected in Image 2 locate (x,, y;). =12...... A 
: Label of displacement vectors. 
ta, reis A (2) 
,' Label of no correspondence. 
Ar (Ar, Ay). Ds Ay SD, —-D<Ay, <D (k#Pp) (3) 
P( ,): Probability that vehicle a, has label ,. We refer to label probability. 
S P(A)=1, O<P(A) SI. (4) 
k 
Kb. 0 =1,2,...... , L) are candidates corresponding to a, a; has L+1 label. 
DE ae we RY (5) 
where 
Aa € A Ag = (yg =X Ya 7 Di) = (Ax, Avy) (6) 
The square of distance of vehicles’ color in the RGB space is 
C( Qm Gyon Gay iG )-g2 0.9) (1 G)- 5203) (7) 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B5. Amsterdam 2000. 279