over 95% has been achieved with a high reliability. The 
positions of detected moving vehicles have been inputted to the 
AdaBoost machine learning algorithm to further improve the 
accuracy of vehicles’ image positions. This is because the 
vehicles’ speed calculation is highly dependent on the accuracy 
of vehicles’ image position. Then, ground positions of each 
detected vehicle from MS-1 and MS-2 images have been 
computed using sensor model (RPC) provided by WorldView-2 
satellite. 
This paper begins by discussing the methodology developed to 
detect moving vehicles. Then, methodology to compute 
vehicles’ information is discussed. Finally, results and 
conclusions are presented. 
2. METHODOLOGY 
In this paper, a Principal Component Analysis (PCA) based 
method has been developed to detect moving vehicles from 
Worldview-2 MS-1 and MS-2 images. The workflow of the 
methodology developed is shown in Figure 1. Then, AdaBoost 
learning algorithm based method has been developed to 
compute vehicles’ information. The work flow of vehicles’ 
information computation is shown in Figure 5. 
2.1 Study Area and Data used 
WorldView-2 imagery of a part of Moncton, a city in New 
Brunswick, Canada, has been used for this study. This 
WorldView-2 image was provided by DigitalGlobe® Inc. to 
Bahram Salehi (University of New Brunswick) through “The 
DigitalGlobe 8-Band Research Challenge" contest. The image 
was taken on October 5, 2010. The WorldView-2 imagery 
includes Pan image MS-1 (BGRNI) image, and MS-2 
(CYREN2) image. The MS-1 and MS-2 bands are stacked 
together as one MS image with 8-bands. 
2.2 Image Resampling 
WorldView-2 MS images have spatial resolution of 2m; 
therefore, small objects like vehicles are not clearly identifiable. 
To make vehicles more identifiable, the MS image has been re- 
sampled to 0.5m using cubic convolution resampling method. 
2.3 PCA Computation 
MS-1 and MS-2 images constitute different spectral 
wavelengths; therefore change detection methods are incapable 
of detecting moving vehicles. In this paper, principal 
components of MS-1 and MS-2 images have been computed. 
As shown in Figure-2 and Figure-3, vehicles are more 
distinguishable in the second principal component. Therefore, 
second principal components of MS-1 image (MS-1: PCA2) 
and MS-2 images (MS-2: PCA2) have been selected for further 
processing. Also, the principal component of MS image, which 
has 8 bands stacked together, has been computed. As shown in 
Figure-4, the vehicles are again more distinguishable in the 
second principal component (MS: PCA2). Furthermore, as 
shown in Figure-4, the second principal component has two 
positions of one moving vehicle. This result is very useful for 
detecting moving vehicles from the MS-1 and MS-2 images. 
Finally, after PCA computations three images, MS-1: PCA2, 
MS-2: PCA2, MS: PCA2, have been selected for further 
processing. 
  
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 
   
  
| WorldView-2: MS image | image Resampling | 
| 
  
  
  
I 
| MS-1:BGRN-1 | |MS-2:CYREN-2| | Ms |] 
| = 
| 
Y 
| PCA Computation | 
l Y Y 
| MS-i:PCA-2 | | MS:PCA-2 | | MS-2:PCA-2 | 
| | 
Y 
Threshold 
| Band Adjustment | 
| 
i J 
| (MS: PCA-2) - (MS2:PCA-2) | | (MS: PCA-2) - (MSI: PCA-2) | 
| i 
Y 
| image Filtering | emt Otsu Threshold 
de Yl 
| MS-1: Moving Vehides | | | MS-2: Moving Vehicles | 
Figure 1: Workflow for vehicle detection 
  
  
  
  
  
  
Threshold 
  
  
  
  
  
  
  
  
  
  
2.4 Band Adjustment 
The histograms of MS-1:PCA2, MS-2:PCA2 and MS: PCA2 
images have been adjusted to a common mean and standard 
deviation. This process has improved the accuracy of moving 
vehicle detection. 
2.5 Moving Vehicle Detection 
After the Band Adjustment, a change detection process has been 
applied to the images to detect moving vehicles. Change 
detection is an important process in remote sensing applications 
(Copping et al., 2004; Tronin, 2006). In the change detection 
process, two images of the same scene captured at different time 
instances are used to detect changes. Therefore, change 
detection can be expressed as: 
Lou at b (1) 
Where 7,; and 7,; are the images captured at time t, and t; and a 
and b are the constant scalar values. The aim of change 
detection process is to model the constants a and b. A variety of 
change detection algorithms are available; however, in this 
study, it has been found that the differencing method for change 
detection is efficient and best suited for detecting vehicles from 
WorldView-2 MS imagery. As shown in Figure-2 and Figure-3, 
the MS-1:PCA2 and MS-2:PCA2 images have different 
positions of a moving vehicle whereas MS: PCA2 image 
(Figure-4) has two different positions of the same moving 
vehicle. Thus, moving vehicles from the MS-1 image have been 
detected using equation (2) and moving vehicles from the MS-2 
image have been detected using equation (3). 
(MS-1 Image)Moving vehicles = (MS: PCA2) — (MS2: PCA2)-T, 
@) 
(MS-2 Image)Moving vehicles = (MS: PCA2) — (MSI: PCA2)-T; 
(3) 
Where parameters T, and T» are the thresholds which are used to 
eliminate outliers appears after differencing process.