r cost and 
ring some 
amera for 
se kinds of 
heat inside 
ptic fibres 
n lighting. 
s less than 
18 
reflectivity 
constraints: 
lge of focal 
ortions and 
1.5 meters 
se a 7.5 mm 
culated mast 
yaw) were 
ent accuracy 
hting 
4. MODELLING SOFTWARE 
The hard environmental constraints and the high amount of 
unknown elements led us towards manual modelling methods. 
To enhance the model elaborated from plans and photos, several 
photogrammetry software were looked upon. They were not 
selected because we were not sure to get enough reliable visual 
indices in the images. The ring-shaped lighting around the 
camera highlights different features as the carrier moves. It 
complicates the detection of homologous points. Moreover, the 
viewpoint is constrained along a vertical line below the access 
hole. 
Integrating geometric constraints is a way to cope with a bad 
viewpoint configuration (McGlone 1995, van den Heuvel 
1998). They can be derived from a semantic approach (Liedtke, 
Grau and Growe 1995, Koehl and Grussenmeyer 2000), but this 
relies on a knowledge base, which is neither available nor 
reliable in our case. They can be explicitly given during the 
reconstruction work (Poulin, Ouimet and Frasson 1998, 
Hrabacek and van den Heuvel 2000). However, the use of a 
CAD modelling system provides a more interactive way to 
derive them (Debevec, Taylor and Malik 1996, Ermes 2000). 
These geometric constraints can also be combined with metric 
knowledge to improve the achieved model reliability (Bürger 
and Busch 2000). 
The interactive modelling software Pyramide (Even and Marcé 
1988) was selected. Pyramide is a 3D modeller originally 
designed at CEA for tele-robotics purposes. According to what 
he sees in a single or multiple images, the operator selects some 
solid primitives (box, cylinder, cones, torus, etc.), and makes 
them match relevant features overlaid on the images. Position 
and dimensions of these primitives are directly controlled with 
the mouse. Assistance based on geometric constraints setting is 
provided to make this work intuitive. Some semi-automatic 
functions using geometric reasoning (Even, Fournier and Gelin 
2000) are proposed to model pipes. Image processing functions 
(Even and Malavaud 2000) help the operator to estimate object 
orientation, position or dimension. 
Earlier uses of Pyramide were for robotics applications to build 
a geometric model of the mobile robot environment. The 
software was also used to localise objects into a robot reference 
frame and to complete a pre-defined model for maintenance 
tasks with a manipulator. In these applications, the problem was 
either to give an accurate position of objects in the scene or to 
provide a rough model of the environment. Recent comparisons 
with a commercial photogrammetry tool revealed the good time 
improvement when modelling with Pyramide for equivalent 
results in terms of accuracy. A major advantage of Pyramide is 
its on-line modelling profile, and its high flexibility to cope with 
unknown and unpredictable situations. 
5. MISSION PREPARATION 
In order to build a model with Pyramide, we need to have some 
correct information on camera position as well as on image 
formation. Focus of the camera was sealed in a position 
allowing having a clear view 1.5m away from the camera. The 
camera was calibrated using a model including distortions. 
Using our camera calibration software, we got nine internal 
parameters representing the optic centre (u0, v0), the focal lens 
(fx, fy) three parameters for radial distortion (al, a2, a3), and 
two parameters for tangential distortion (pl, p2). The external 
parameters (r11, 133) for the rotation and (Tx, Ty, Tz) for the 
translation allow setting a reference on the camera support. 
plas KAHN VE LT (1) 
u=u + ——— are vor Wf T 
dx EX 5th tl 
v=v+ gra dow roe mom Ie (2) 
dy aX tin via +1z 
With 
dOxr = (u —u0)dx(a,r” tac a,r°) (3) 
dou = (v vo)ay(a 7” ra, tay) (4) 
And 
dOxt = p, (^ +2(u —u0) dx’ J+ 2 p, (u — u0)dx(v — »0dy (5) 
dOyt = p, (^ +2(v— v0)’ dy’ }+ 2p,(u — u0)dx(v — vo)dy (©) 
r, the radial distance from main point valuing: 
r 2 A(u -u0) dx? * (v - voy ay? (7) 
  
(dx, dy) being the size of a pixel and fx = f and fy- I 
dx dy 
(u, v) are the coordinate of the point in the image. 
The camera, ring lighting were mounted tight to the carrier. 
Several tries were then conducted off-line to check the global 
modelling process with Pyramide, the camera calibration, the 
carrier accuracy and linearity of its joints. As a remark, we did 
not need to ask the carrier to move to accurate positions but 
rather to be able to measure its actual position with a great 
accuracy. Therefore, the arm carrier arm was not calibrated. 
This preparation stage was essential before modelling since 
there was no question to get the camera back once it had been 
installed in the room. 
6. INSTALLATION OF THE EQUIPMENT 
6.1 Remarks on the carrier 
Because of arm architecture, reachable viewpoints all laid along 
a vertical line. Only triangulation within vertical planes could be 
used. The possibility to introduce geometric hypothesis within 
the modelling process was very interesting to compensate for 
these constrained viewpoints. 
The structure of the carrier (3 degrees of freedom only), the 
resolution of the sensor, and the impossibility to choose other 
viewpoints than along one vertical axis, did not allow making 
the distinction between objects (little pipes, valves) far from the 
camera. 
6.2 Remarks on the obstacles 
After the introduction of the arm carrier in the hole, a first set of 
images was captured. It appeared during this first operation that 
we could not avoid the carrier to touch the environment in most 
configurations. An obstacle, the heat exchanger, was in the way 
20 cm under the ceiling. Therefore, the translation movement 
was not linear when the arm was in contact to the obstacle. We 
suspected even the translation axis to change orientation during 
pitch movements. Measures pointed out a 2.5 degrees deviation 
relatively to the vertical axis. 
These contacts between the carrier and the environment 
seriously changed the conditions of the experimentation. Joint 
position of the arm could not be trusted. Positions when the arm 
—177—