International Archives of the Photogrammetry, Remote Sensing
Of course, in an overall system design, measurements provided
by odometers and GPS can help the image processing by reduc-
ing the search spaces and as consequence by reducing processing
times and robusteness.
I F m WA Ss. ni z E =
MMS image fie = DNR MMS image
® i |! Image and template | Image and template (edo)
| Ty marching by dynamic}. matching by dynamic
|
v
programming | programming Ï
Y: Je Y
2D line set (tdt) |
2D line set (0 DFM (0) | | DEM (tdt)
| NN |
| 3D plane extraction (1) | 3D plane extraction (t+dt)
Vanishing points | ; vs Ÿ ; Rad idi — | Vanishing points |
detection (t) | Facade i | Facade | | | detection (t«dt)
J _ orthoimage (J | orthoimage (+d) J (50 | ;
|
Absolute pitch and | Matching 3D planes by FFT | Absolute pitch and |
roll angles (t) template correlation roll angles (tdt) |
|
| ^ E . . Ÿ . TN
I | Estimating relative translation
i and rotation (t, t+dt)
Relative and partial absolute
pose estimation
Figure 2: Georeferencing strategy of terrestrial images using fea-
tures.
3.1 Measuring relative pose from the image sequences
Exterior orientation in photogrammetry and pose estimation as it
is mostly called in computer vision is a popular research subject.
In this paper, we investigate the use of higher-level geometric fea-
tures such as 3D points, 3D lines or 3D planes generated from a
range measurements unit as observed geometric entities to im-
prove the automation of the sensor pose estimation.
The cameras used on our system have fixed focal lengths. They
are calibrated on a 3D target polygon and on a planar textured
wall. Focal length, principal point of autocollimation, principal
point of symmetry and the coefficients of a radial distorsion poly-
nom are estimated. The image residues are generally of a tenth
of a pixel. The relative pose between the different cameras com-
posing the rig are also determined on the 3D target polygon.
As the cameras are perfectly synchronised and since the relative
orientation of the cameras on the vehicle are known, all the cam-
eras and their very different viewing geometry will contribute to
determine robustly and accurately the vertical and horizontal van-
ishing points and as a consequence to estimate the roll and pitch
angles of the platform (with respect to the vehicle displacement).
Figure 3: Estimating the roll and the pitch angles of the platform.
3.22 Sub-Pixel Features Detection
The first step of the straight lines detection stage is the computa-
tion of image derivatives followed by the non-maximum suppres-
sion using the optimal Canny-Deriche edge detector (Deriche,
1987). The contour pixels are thereafter chained and are subpixel-
localised by finding the maxima of an analytical function fitted
and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
Figure 4: MMS Images taken by use of the vertical basclines.
through the sampled gradient measurement in the gradient direc-
tion. This improvement in localisation reduces in a significant
way the aliasing affects thus the shapes are much smoothly de-
scribed and as a consequence determining "intelligent" thresh-
olds for polygonal approximation is much easier. The estimation
uses an iterative merging process based on the maximum residual
using the orthogonal regression (Taillandier and Deriche, 2002).
One of the advantages of using orthogonal regression is that the
errors associated with the straight lines parameters can be de-
termined. First, the polylines whose merging gives a minimal
maximum residual are merged. The tolerance on the polygonal
approximation acts us to stop the process when the merging has
a maximum residual above a threshold given by the user. Once
the polygonal approximation is done, the parameters 0 and p of
the lines underlying the segments as well as the variance covari-
ance matrix of these parameters are estimated by using the results
of (Deriche et al., 1992) algorithm and under the assumption that
the edges detected by the Canny-Deriche detector have a variance
given by :
var — : 9
ji 0 2
where o can be determined through the ratio signal/noise in the
images (see Figure 5).
Figure 6: Illustration of detected line segments.
International Arch
3.3 Vanishing P
Due to the effect:
parallel in the 3D
pending on the or;
be finite or infinit
are many method:
A good review is
Here the vanishin;
the method propo
multi-head viewin,
segments of each i
corresponds a 3D
on the same plane
horizontally orient
inant groups of co
duce maxima on |
each segment (A a
uncertainty bound
tion to take into a
segment extractior
to guide the searc
tion thus indirectly
weigh more than s
indicates the degre
(a)
Figure 7: (a) Pla
ing the Gaussian s
space. (c) Bounds
the tangent plane a
Figure .8 shows th
vertical baseline h
ometric reasons, tt
mixing images fr
when the full syste
(a)
Figure 8: (a) & (
each MMS image
Multi-view vertical
Once the vertical v
the non vertical seg
ing equator (consi
equator) to find all
accumulator cells c
Each vanishing poi
tation relatively to
