of not only 
| x-y domain. 
Ss“ (a, b) of 
ime £. When 
between the 
le. 
(9) 
(10) 
clusters are 
od to aerial 
> as follows: 
8bit); 
ixels. 
| Was used to 
arameters in 
e specified. 
for velocity 
hreshold for 
aracteristics. 
v, H». 9s: 
meters w in 
Or, since an 
'er, we used 
  
  
  
  
  
Figure 4. Result by spatio-temporal clustering method 
Figure 4 shows a part of the result by the proposed method. In 
this image the velocities of vehicles are about from 0 kim/h to 
50 km/h. Recognized vehicles are circumscribed by rectangles. 
Only on the road is set as area of interest. In this experiment all 
vehicles in 77 vehicles, whose movements include direct 
advances, left and right-hand turns, lane changes and stoppage, 
could be recognized (recognition rate is 10094). 
4.) Application to Different Spatial and Temporal 
Resolution Image 
The more spatial and temporal resolution is low, the more data 
quantity decrease. Consequently, the load of observation and 
telecommunication equipments is reduced. Therefore, it is 
important to grasp spatial and temporal resolution influence for 
vehicle manoeuvres recognition. We also applied the proposed 
method to different spatial and temporal resolution images as 
follows: 
spatial resolution: — 10, 20, 30, 40, 50, 60, 70 (cm); 
temporal resolution: 1/30, 1/10, 1/5, 1/2, 1 (seconds). 
Here, temporal resolution means a time interval of successive 
image. The data of the image is same as the previous 
experiment. 
Figure 5 depicts the vehicle recognition rates with various 
resolutions images. The proposed method was more sensitive 
to temporal resolution than to spatial resolution. In the case of 
under 40 cm or 1/5 seconds, over 80% of recognition rates 
could be obtained. With probabilistic  relaxation-based 
approach, the sensitivity to temporal resolution could be 
alleviated. 
4.3 Evaluation of Vehicle Position Accuracy 
In practice, vehicle position is employed as data for application 
to traffic engineering and planning. Moreover, velocity or 
acceleration calculated from the position is employed. Hence it 
is very important to grasp of the vehicle position accuracy. 
The vehicle position accuracy was evaluated thorough 
comparison between results by the proposed method and 
—565— 
temporal resolution (s) 
  
spatial resolution (cm) 
Figure 5. Examination of various resolutions 
x-coodinate (pixels) 
400 . 300 2200 -100— 
  
   
  
  
  
  
  
  
  
  
  
  
  
_ Fai Dm — -  -100 
% vehicle trajectories] 8 
é 2 200 & 
/ 5 
| à 
300 = 
LH 
| 400 
-500 
  
—s— proposed method 
— manual 
Figure 6. Examination of vehicle position accuracy