International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B5. Istanbul 2004 
RANSAC estimation of corresponding matches yield a more 
reliable recovery of the camera structure. 
Moreover, if the left and right shots are simultaneous, the 
epipolar geometry between the images is unchanged through 
the whole sequence: a first calibration of the relative camera 
leads to a precise and reliable computation of the fundamental 
matrix that can be used to “robustify” the matching in a guided 
procedure. As relative orientation and baselenght are constant 
along the sequence, the Euclidean reconstruction achieved 
using this information, allows to join different image subsets 
and eliminates the scale factor ambiguity. : 
Merging the two approaches assures a great improvement in the 
procedure performance: while the "along sequence' pair leads 
to a great number of correspondences (most of which correct), 
the epipolar constraint arising from the *across sequence' pair 
tends to eliminate some ambiguity (i.e. points on the road 
markings) or troublesome wrong matches (i.e. building facade 
texture). This merging is performed through a trifocal tensor 
estimation using a robust algorithm. The resultant sub-block 
(using two consecutive trifocal tensor computation we obtain a 
symmetric four-image block) has a balanced size: the distance 
between consecutive frames is approximately 4 m, while the 
baseline of a synchronous pair is about 1.7 m. Since the camera 
pose estimation is obtained within this symmetric configuration, 
matched points can be found also in the vicinity of the 
projection centres (so filling more uniformly the image format); 
using two trifocal tensors provides a strong filter and leads to a 
better constrained and redundant bundle block adjustment at the 
end of the pipeline; finally, though less precise, points far away 
tends to be tracked along many images of the sequence, 
providing ties between sub-blocks. 
4. FIRST TESTS AND RESULTS 
In this chapter we present a summary of the results achieved 
during the test procedure. Although the testing of the 
implementation cannot be yet considered completed, we 
processed a fairly representative set images, in terms of 
arrangements of the images in the processing sequence, in terms 
of road traffic, vehicle speed and scene background. Given the 
variety of situations along roads, no definitive conclusion can 
be drawn yet, but tests have been useful to understand how and 
when satisfactory results can be obtained using the proposed 
approach. In the first part of this section we illustrate results on 
a block along a small countryside road around Parma, as 
showed in figure 4; then we evaluate through simulated data set 
the error propagation along a sequence of about 250 m and how 
points or camera constraints can improve the motion and 
structure computation. 
4.1 Matching procedure and metric reconstruction 
As already said all the tests were performed using two digital 
cameras (Basler A101f) with a 8 mm focal length and a 
resolution of 1300x1030 pixel. Since the camera lenses we use 
produce a strong barrel distortion effect on the images, we first 
determined their calibration parameters with a build-up test 
field. The estimated distortion model is correct up to 0.5 pixel. 
A good camera distortion model is necessary because 
attempting an automatic self calibration never gave the desired 
and expected results. 
About 2000 feature points have been extracted in every image 
of the sequence using the Harris operator: the number of feature 
to accept was determined, considering the camera resolution 
and quality, by finding a compromise: using too many points 
the epipolar and trifocal estimation may leads to uncertain 
results, whenever they are too close to each other so that an 
ambiguous matching arise; on the contrary with fewer points 
their ground distribution is poor and the camera pose is affected 
by a weak geometry. 
Through the disparity threshold a first putative matching is 
computed between the images: though the a/ong sequence 
approach tends to give more correspondences than the across 
sequence one, the difference is negligible (see table 5). 
The main difference between the two approaches arise during 
the first outlier filtering: here, the perspective differences 
between left and right images leads to a more difficult matching 
between the putative correspondences; with a least squares 
matching approach, the differences arising from the disparity in 
the viewing angle may be taken in account and more matches 
might be obtained. 
In order to limit computation time for the across sequences the 
epipolar geometry is evaluated only on the first pair using a 
robust algorithm; then, in the other pairs, a guided matching 
procedure, using the estimated fundamental matrix between left 
and right images, is performed. 
  
  
  
  
  
Figure 1. Matched points between left and right image of a 
synchronous stereo pair (across sequence). 
  
  
  
  
  
Figure 2. Matched points between consecutive left images 
(along sequence). 
The algorithm proceeds joining two different epipolar matched 
data sets in order to perform the trifocal outlier filtering; the 
common points are therefore fewer than those found separately 
in the image pairs. 
The trifocal tensor manage to eliminate those outlier that 
satisfied the epipolar geometry constraints: in figure 3 we can 
see how the parallax effect of the branches of the tree on the 
background of the white building, arising from the different 
standpoint of the left and right cameras, satisfies even the 
epipolar geometry but is spotted by the tensor geometry. 
  
me rn 
abi 
Re-— ^ £M ad 
  
  
be