The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Voi. XXXVII. Part Bl. Beijing 2008 
6 
1.2 Outline 
The following chapter presents an example implementation of a 
multi-camera VIDS. The utilized processing chain is briefly 
presented in order to emphasize the need for precise knowledge 
of the exterior orientation. Chapter 3 introduces the different 
implemented approaches to determining the exterior orientation 
of the cameras. This is followed in chapter 4 by an evaluation 
of the approaches concerning their accuracy and usability in 
respect to the given task. The last chapter summarises and 
concludes the results of the evaluation. 
2. APPROACH FOR A MULTI-CAMERA VIDS 
2.1 Processing Approach 
A multi-camera setup has been installed using three cameras to 
observe the traffic intersection Rudower Chaussee/ 
Wegedomstrasse, Berlin (Germany). The cameras cover 
overlaid or adjacent observation areas. Thus, the same road user 
can be observed by different cameras from different positions 
and angles (Figure 1). Using image processing methods the 
objects of interest can be found in the image data. 
I 
Figure 1. Visualisation of the multi-camera-setup 
In order to enable the tracking and fusion of detected objects in 
the observation area the image coordinates of these objects are 
converted into a common world coordinate system. In case of 
poor quality of the orientation parameters, the same objects are 
observed from different positions. To avoid misidentification of 
objects derived from different cameras, a high precision 
transformation of their image coordinates into the object space 
coordinates is required. Therefore, a very exact calibration 
(interior orientation) as well as knowledge of the position and 
the view direction (exterior orientation) of the cameras is 
necessary. 
The approach presented here can be separated into three main 
steps. Firstly, all moving objects have to be extracted from each 
frame of the video sequence. Next, these traffic objects have to 
be projected onto a geo-referenced world plane. Afterwards, 
these objects are tracked and associated to trajectories. This can 
be utilized to assess comprehensive traffic parameters and to 
characterize trajectories of individual traffic participants. These 
steps are described more precisely below. 
2.2 Video Acquisition and Object Detection 
In order to receive reliable and reproducible results, compact 
digital industrial cameras with standard interfaces and protocols 
(IEEE 1394) are used. 
To extract moving objects from an image sequence the image 
processing library OpenCV was utilized. The algorithm is based 
on a background estimator, which adapts to the variable 
background and extracts the desired traffic relevant objects. The 
extracted objects are then grouped using a cluster analysis 
combined with additional filters to avoid object splitting by 
infrastructure at intersections and roads. 
The dedicated image coordinates as well as additional 
parameters like area, volume, color and compactness can be 
computed for each extracted traffic object. 
2.3 Coordinate Transformation and Camera Calibration 
The employed tracking concept is based on extracted objects, 
which are geo-referenced to a world coordinate system. This 
concept allows the integration or fusion of additional data 
sources as long as their observations can be transferred to the 
same coordinate system. 
Therefore, a transformation between image and world 
coordinates is necessary for a multi-camera system. Using the 
collinearity equations (1), the world coordinates X, Y, Z can be 
derived from the image coordinates x', y': 
r u (x'-x 0 ) + r 23 (y'-y 0 )- rii c (1) 
v_v , /7 7 ^ r u (x'-X 0 ) + r 22 (y'-y 0 )-r i2 C 
1 ~ J 0 + ¿0 ) , , . . 
r x2 (x-x 0 ) + r 22 (y-y Q )-r 22 c 
where X, Y = world coordinates (to be calculated) 
Z = Z-component in world coordinates (to be known) 
X 0 , Y 0 , Z 0 = position of the perspective center in 
world coordinates 
r n , r 12 ,..., r 33 = elements of the rotation matrix 
x', y' = uncorrected image coordinates 
xo, yo = coordinates of the principle point 
c = focal length 
The Z-component in world coordinates can be deduced by 
appointing a dedicated ground plane. Additional needed input 
parameters are the interior and exterior orientation of the 
camera. The interior orientation (principal point, focal length 
and additional camera distortion) can be determined using a 
well known lab test field. The 10 parameter Brown camera 
model was used for describing the interior orientation (Brown, 
1971). The parameters can be determined by a bundle block 
adjustment as described in (Remondino and Fraser, 2006). 
In order to calculating the exterior orientation of a camera, 
hence determining its location and orientation in a well known 
world coordinate system, different approaches can be applied. 
An important set of these approaches are presented and 
evaluated in the following chapters. 
2.4 Tracking and Trajectory Creation 
The tracking algorithm is supposed to provide object data 
information combined in a so-called state vector with respect to 
time. The state of an object can be described as position, 
velocity and acceleration in X-, Y- and Z-direction. Features 
like form, size and color can be added. The first task is the 
object identification in a video sequence by its predicted state