7A-4-2 
Transportation 
Object 
Imaging and 
GPS/INS 
Georeferencing 
Figure 2. Overview of the approach 
• Back projection of objects from a database onto images 
• Verification of the existence of objects 
• Multinocular line reconstruction for accurate object 
positioning 
The mobile mapping image sequences collected have been 
georeferenced using GPS/INS positioning and photogrammetric 
CCD calibration techniques. Consequently, the orientation and 
position parameters for each camera exposure station are determined 
with respect to a global coordinate system. For more detailed 
information on image georeferencing methods, one can refer to 
Schwarz and El-Sheimy (1996) and El-Sheimy (1996). 
3. VERIFICATION OF THE EXISTENCE OF 
OBJECTS 
The mounting of the stereo cameras in the VISAT system 
follows the way that the baseline is constrained approximately 
to be parallel to the ground plane and the camera roll angle is 
configured to be very small. As a result, the projections of 3-D 
line features of the objects in images are also approximately 
vertical. 
3.1 Detection of Vertical Edges 
J 
Edge magnitude and edge direction are two main components of an 
image edge element. We firstly recognize from experiments that 
information related to edge direction is valuable since the edge 
magnitude is easily corrupted by noise and physical factors, such as 
shadows. Therefore, the computed edge direction is used as the main 
evidence for inference. In the following, a three-step algorithm to 
determine vertical edges is described: (a) determination of the 
existing region of possible edges; (b) determination of the existing 
region of thinned edges; and (c) determination of the existing region 
of vertically oriented edges. 
3.1.1 Determination of the Existing Region of Possible Edges 
We start with convolving an input image using the six directional 
masks (Tao, 1997). After convolving the masks with the original 
image, the magnitude of the convolved output and the direction of 
the mask giving the highest output at each pixel are recorded as edge 
data. A very small threshold is set in order that all possible edges are 
detected in the edge existing region. The direction value is encoded 
as 0 - 5 corresponding to the six directions previously defined. An 
example of the obtained edge existing region is shown in Figure 3b. 
3.1.2 Determination of the Existing Region of Thinned Edges 
The thinning algorithm is designed as follows, where the obtained 
direction values of edges are used to constrain the thinning 
procedure: an edge element exists if the following two conditions are 
satisfied 
The transportation objects to be examined are firstly queried from an 
existing database. The position of the desired object is projected onto 
the corresponding images using photogrammetric collinear 
equations. A global coordinate system, i.e., UTM coordinate system, 
is used during the projection. The normal speed of the mobile 
mapping platform (van) is around 50-60 km per hour and the image 
capture rate is about 0.4 second. As a result, an object is usually 
visible in 3-4 image stereo pairs. 
Due to the use of the prior information of the object position, the 
approximate position of the object in the images is determined. Thus, 
the processing window of the object in the images can be predicated 
(shown in Figure 2). It is recognized that the use of this prior 
information greatly simplifies the procedure of object detection, and 
moreover, makes the processing results much reliable. 
The module of verification of the existence of the objects is to 
identify whether or not the object exists (or misses). It involves a 
number of processing steps: detection of vertical edges, formation of 
line segments and feature correspondence of line segments. Once the 
existence of the object is verified, the module of multinocular line 
reconstruction is executed to determine the accurate position of the 
object, and then update the position information of the object in its 
database. 
• the edge directions of the two horizontally neighboring pixels 
are within one unit (45°) of that of the center pixel; and 
• the output edge magnitude at the pixel is larger than the edge 
magnitudes of its two neighbors in a direction normal to the 
direction of this edge. 
If the above two conditions are met, the two neighboring pixels are 
disqualified from being candidates for edges. An example of edge 
thinning is given in Figure 3c. 
3.1.3 Determination of the Existing Region of Vertically 
Oriented Edges 
A further refinement is conducted using the accurate values of edge 
direction. In this step, we use the Sobel gradient operator to compute 
the direction values of the existing edges. Since a large amount of the 
undesired edges have been screened out through the previous two 
steps, the computation of edge gradients is only performed on a 
small set of edges, and the savings on the computation time are 
significant. After the computation of gradients, a rigorous tolerance 
of direction value, 15°, is applied to filter out all unqualified edges. 
The processing result after this step is illustrated in Figure 3d.