EVEN, Philippe
time and restricts the camera use to static views. Moreover it inserts some blur in the final image. In order to ensure a
correct match between the model and the images, the camera viewpoint must be computed from the mobile robot relative
localization module (Even et al., 1989) or from the manipulator arm joint values (Even et al., 1999). Pyramide interest
has been verified through many experimental campaigns in realistic conditions.
Time is the most important limitation factor for this technique. Therefore most of the works performed these last years deal
with solutions to speed up the interactive modelling process. Structural knowledge on particular environments has been
integrated into specialized modules, based on dedicated primitives handled through an optimized modelling operating
mode and automatic constraints management. A piping module has been developed and evaluated (Even et al., 2000).
Considerable time is spared for the modelling of piping elements. The remaining objects are modeled using the generic
module. A large improvement is also obtained by setting the description level to the mission needs. The functional
classification proposed in (Even et al., 1999) shares the objects into one of the three following types : out of reach objects
are coarsely modelled, just letting some distinct features for navigation purposes; potential obstacles are composed of
surrounding volumes, their accuracy being directly connected to the security distance of a collision avoidance module;
manipulated objects are more detailed with regards to the required geometric features. The basic idea is to provide the
operator with some task-oriented requirements in order to avoid useless model refinements.
A large enhancement can also be expected from the integration of automatic assistances based on computer vision tech-
niques. For instance the automatic fitting of a primitive on image features has been integrated into several modelling
systems (Hsieh, 1995, Debevec et al., 1996, Kim, 1999, Giilch et al., 1999). The operator very quickly provides an coarse
initial solution. Manual fine matching is tiring and takes a lot of time. This task is autonomously performed using an opti-
misation algorithm based on the minimisation of a distance between the drawn segment and the extracted edge segments
in the image. Pyramide has also been improved by integrating such primitive fitting in semi-automatic mode (Bonneau
and Even, 1993).
3 EDGE SEGMENT SPECIFICATION PURPOSE
The basic modelling principle in Pyramide consists in matching solid primitives on the relevant image features. However
the fine control of an object orientation in space using a 2-D input device such as a mouse and only visual feedbacks is
a difficult task. Computing the object orientation from image features can help a lot. Once the object is well oriented,
its matching is easily obtained with the mouse. Pyramide provides several methods to compute an orientation from the
available edge segments in the images (fig. 2).
A) Parallel lines B) Single line C) Revolution axis D) Coplanar features E) Orthogonal lines
Figure 2: Available methods for the extraction of 3-D features from edge segments specification.
e A) Extraction of a vector from the projection of parallel lines in one view.
The intersection of the edge segments is the vanishing point associated to the 3-D lines common direction. It is used
to compute the lines unit vector.For a better accuracy, wide angle lens and distant edge segments in the image are
recommended (Shufelt, 1999b).
e B) Extraction of a line from its projection in two views.
For each view, the edge segment defines an interpretation plane passing through the projection center. The 3-D line
lies at the intersection of both edge segments interpretation planes. This method is used to align an object on a given
edge. The object is then moved along and rotated around its edge. Accuracy requirements deal with the selected
edge segment length, with its sharpness and with the angle between the interpretation planes.
t3
N
A
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B5. Amsterdam 2000.
All
toa
try 1
Edg
Acc
autc
ima
sol
seg
seg
Sev
attr.
sho
dep
mo
The
sibl
assi
mo