ET
NA D Se Et o TRA Tac i eA aC imd joo uc IE Bia a ares
PT
180
Parameters a; and a4are especially introduced for digital cameras, whose pixel ccordinate system often have a signifi-
cant scale difference and non-orthogonality.
Additional camera parameters can be introduced, e.g. for tangential distortion. But for this application the parameter
set mentioned above gives image point coordinates with an accuracy better than the matching accuracy and is there-
fore sufficient. :
For the calibration of the Kodak DCS 200 a flat point field with 500-1500 circular target points is used. A set of 12-20
images from different directions (cf. Figure 5) is taken and the measurements are done for at least four of the points in
each image semiautomatically; the rest of the points are measured automatically by least squares template matching.
The approximate values result from an 8-parameter transformation which is updated after the measurement of a
single point. It is possible to apply the whole calibration measurement fully automatically, if at least four points are
identified and are measured automatically.
3.3 Surface reconstruction
After the reconstruction of the image orientation the system is ready to start with the automatic surface measurement.
Two measurement options have been realized so far. The interactive input during the preparation step is kept as small
as possible and comprises the definition of the measuring area, the digitization of exclusion areas, borderlines, starting
points, and control parameter.
3.3.1 Point cloud.generator
This measurement option of the system is mainly based on an approach which has been originally designed for auto-
matic DEM generation (Krzystek, 1991). This approach takes advantage of many arbitrarily distributed 3D points which
are found by means of feature-based or area-based matching techniques to provide a dense surface description. In
order to overcome the restrictions of the topographic application field, it has been spezialized for close range applica-
tions. The intention of the modification was to contribute to new trends in car design by a point cloud generator which
digitizes surfaces very densely. The points of the cloud are structered in grid lines defining facets or polyhedrons. The
orientation of the grid lines is determined by a reference plane which is to be defined in advance.
Basically, the approach is characterized by the feature-based matching technique being hierarchically applied in image
pyramids and a robust surface reconstruction using finite elements. Measured 3D points, together with curvature and
torsion constraints are introduced as observations. Since the observations are weighted down in a two-step iterative
least squares adjustment, the algorithm is principally suited to preserve surface edges like characterlines. The spacing
of the grid points is by default relatively small (e.g. 15 pixels) and is automatically adapted to the local surface curva-
ture. Because of the hierarchical approach through image pyramids no initial surface is required to start the surface
reconstruction.
Actually the key idea of the point cloud generator is the automatic measurement of a large number of irregularly dis-
tributed 3D points. Let us presume a typical stereo camera configuration with a camera-to-object distance of 1.5 m
and a base-to-height ratio of 1/3. The camera in use should be a standard middle format camera with a focal length of
10 cm, and the pixel size of the digital image should be 15 um, meaning that the pixel size on the object is in the order
of 0.2 mm. Since the matched feature points usually have a precision of 1/3 pixel, the internal measuring accuracy of a
single 3D point in object space is expected to be V2 *1/3 pixel *15*3 = 0.31 mm. If the object is for instance covered
by an artificially created texture (e.g. projected random pattern), typically 30 to 40 measured 3D points are found in a
single finite element mesh with a size of 15 pixel x 15 pixel = 3.4 mm x 3.4 mm. Because of this high redundancy, the
- points of the finite element get a higher accuracy than the single matched 3D points, which is for this particular case
in the order of + 0.09 mm. Nevertheless, the individual matched 3D points can easily be refined by least squares
matching techniques, which is reasonable in case of smaller finite elements and the presence of poor texture, respec-
tively. With regard to the camera configuration mentioned above, a point cloud covering an object area of about 1 m x
1 m with approx. 100000 3D points on average is generated. Figure 2 shows a point cloud created on a concept
model.
3.3.2 Profile measurement
Especially in car industry, the digitization of design models is required along profiles which are defined by the intersec-
tion of pre-defined planes with the surface. The intersection planes are usually parallel to the car coordinate system. In
some cases special planes are to be defined perpendicular to 3D polylines or are rotating along a straight line as so-
called radial planes.
The measuring strategy in this second option incorporates hierarchically applied area-based and feature-based
matching and guarantees that the surface is measured with profiles wherever possible. Problem areas caused by poor
texture or reflections, in which matching techniques fail, are circumvented in order to keep gaps as small as possible.
Only one starting point has to be provided in advance, whose digitization accuracy must fall in the pull-in range of the
IAPRS, Vol. 30, Part 5W1, ISPRS Intercommission Workshop "From Pixels to Sequences", Zurich, March 22-24 1995
< f^ —-
-— ("7 — 6
m e
