vehicle detection process. The process resulted in two black and 
white images containing moving vehicles of the MS-1 and MS- 
2 images. The vehicle detection results from the MS-1 image 
are compared with manually counted vehicles in Table 1. There 
were 424 moving vehicles in the MS-1 image, and 399 vehicles 
were detected correctly. Only 18 vehicles were missed and 43 
detected vehicles were noise. As shown in Table 1, the producer 
accuracy was 95 % and user accuracy was 90%. The results of 
vehicle detection from MS-2 images are similar to vehicle 
detection in the MS-1 image. Almost all the vehicles of MS-1 
and MS-2 images were detected. The main advantage of the 
developed approach is that there is no need of road extraction 
prior to the vehicle detection. Furthermore, both light and dark 
colour vehicles are detected. A few false vehicles are detected 
due to vehicles’ shadows and lines on the road. 
Table 1: Accuracy of vehicle detection from WorldView-2 
imagery 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B7, 2012 
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
Vehicle Pu Correct | Omission | Commission 
extracted | count 
Roadl | 135 130 125 S 10 
Road2 | 127 119 119 0 8 
Road3 | 76 70 65 5 11 
Road4 | 86 80 72 8 14 
Total | 424 399 381 18 43 
Producer accuracy User accuracy 
(correct/true)x 100 (correct/extracted)x100 
95.49 % 89.86 % 
  
2.8 Speed Computation from WorldView-2 Imagery 
Due to slight time delay in WorldView-2 MS-1 and MS-2 
sensors, the speed of a moving vehicle can be computed using 
vehicle's positions in MS-1 and MS-2 images. The time delay 
between MS-1 and MS-2 is 0.22 second (Tao and Yu, 2011). 
The accuracy of a vehicle's speed computation is highly 
dependent on the accuracy of a vehicle's position determination 
due to the very small time delay. 
The vehicles detected in the previous section have been used to 
detect vehicles’ accurate position on the ground. In the previous 
section, moving vehicles are detected after the resampling of the 
MS image; therefore, if vehicles' centre positions are calculated 
from the detected vehicles, the error can reach up to 0.5 pixels. 
This will directly affect the accuracy of vehicles! speed 
calculation. Thus, to minimize the error in vehicles’ speed 
computation due to resampling, adaptive boosting (AdaBoost) 
machine learning algorithm has been used to detect vehicles 
from the original (without resampling image) MS-1 and MS-2 
images again. To make detection process efficient and accurate, 
first, the approximate positions of moving vehicles are 
computed using previously detected vehicles. Then, these 
positions are inputted to the AdaBoost algorithm to limit the 
search space for the vehicles. The workflow for vehicle's speed 
computation is shown in Figure 7. 
AdaBoost or boosting is a machine learning algorithm which 
builds a strong classifier by linear weighted combinations of 
weak classifiers. A group of boosting algorithms have been 
discussed in (Freund and Schpire, 1997; Freund and Schpire, 
1999). Leitloff and Hinz (2010) have used Gentle AdaBoost 
algorithm (Friedman et al, 2000) to detect vehicles from the 
QuickBird Pan image. They have found that Gentle AdaBoost 
algorithm is most suitable for working with satellite imagery as 
this algorithm is less sensitive to the errors in the training set. 
The Gentle AdaBoost classification algorithm needs a proper 
  
   
  
MS-1 & MS-2 Vehicle's 
image positions 
(previously detected) 
| Training |——»| Gentle AdaBoost Algorithm | 
1 
Me 
Vehicles from MS-1 
| MS-1 Image | | MS-2 image | 
  
  
  
  
  
  
  
Y 
Vehicles from MS-2 
  
  
  
  
  
image image 
Y Y 
Vehicle's center position Vehicle's center position 
in MS-1 image in MS-2 image 
  
  
" 
Vehicle's ground 
positions (X,Y) using 
Sensor Model (RPC) 
[ 
Vehicle's ground 
positions (X, Y,) using 
Sensor Model (RPC) 
I 
  
  
  
  
  
Distance: Ad=vA XZ + AYZ 
Ad —; ÂX 
Vehicle'sSpeed-z- and Vehicle's Direction 8-tan : AG 
  
  
  
Figure 7: Workflow for vehicle’s speed computation 
  
    
Figure 8: Test vehicles on Pan imagery for speed computation 
training to optimize the boosting parameters. In the Leitloff and 
Hinz (2010) approach, after the training process, Gentle 
AdaBoost algorithm has been applied to detect vehicles from 
the roads areas of the Pan image. In this paper, the search space