El-Hakim, Sabry
3.2 Accuracy Verification
In the tests shown in figure 7.A and C, several distances were measured between 3D points. The differences between
the computed distances and the directly measured distances were computed. Also radius computed from surface fitting
of the globe and a circle on the rim in figure 7.B are compared to directly measured values. Deviations from plane in the
walls showing in figure 7.D were also computed. Table 1 displays the results. Accuracy is estimated to be from 1: 2300
to 1: 6000, which is good considering that natural features were used for image registration and point measurement.
Description Actual value | Difference Relative to site size
Distances dl, figure 6-A (5 m view) 1215 -1.5 1: 3700
d2, figure 6-A 3810 -1.9 1: 2900
length of bench, figure 6-C (3 m view) | 2001 -0.5 1: 6000
width of bench, figure 6-C 300 1:3 1: 2300
between targets 13-14, figure 6-C 802 0.5 1: 6000
Fitted surfaces Sphere radius in figure 6-B (2 m view) | 300.5 0.35 1: 5700
Circle (rim) radius in figure 6-B 351 0.85 1: 2300
Planes in figure 6-D (2.5 m view) plane 0.60 offplane | 1: 4200
Table 1: Results of accuracy tests. The values and differences are in mm.
4 CONCLUSIONS AND FUTURE WORK
The described system exhibits notable improvement in flexibility, accuracy, and completeness over existing approaches.
The system is mostly interactive with easy to use interface. Depending on the type of surface and the environment,
certain components are automatic. The main advantage of the approach is its flexibility in that it can use image-based
modeling from multiple or single images, combine multiple image sets, use data from positioning devices, and
integrates data from range sensors such as laser scanners. The accuracy achieved by applying a complete camera model
and simultaneous photogrammetric global adjustment of bundles is sufficient for most applications.
Although this interactive system can be used to model a wide spectrum of objects and sites, it is still desirable to reduce
human intervention, particularly when using a large number of images. Automation is particularly needed for:
e image acquisition and view planning (incremental on-site modeling may be needed),
e point extraction and matching before registration, especially for widely spaced camera positions, and
e determining point connectivity by segmentation of 3D points into groups.
Occlusions and variations in illumination between images affect existing automatic methods for correspondence and
image registration. Therefore, they need images taken at close intervals, which result in too many images as well as
reduced geometric accuracy. In addition, the resulting 3D points are not likely to be suitable for modeling. Therefore,
improved automated methods that do not suffer from these shortcomings are the subject of future research.
ACKNOWLEDGMENTS
I would like to thank my colleagues Angelo Beraldin and Luc Cornouyer for providing the range sensor data.
REFERNENCES
Baillard, C., Zisserman, A., 1999. *Automatic reconstruction of piecewise planar models from multiple views", IEEE
Conference on Computer Vision and Pattern Recognition, pages 559-565.
Beraldin, J.-A., Blais, F., Cornouyer, L., Rioux, M., El-Hakim, S.F., 1999. “3D imaging for rapid response on remote
sites”, SIGGRAPH'99 Technical Sketches, pp. 225.
Chapman, D., Deacon, A., 1998. "Panoramic imaging and virtual reality - filling the gaps between the lines", ISPRS
Journal for Photogrammetry and Remote Sensing, 53(6), pp. 311-319, December.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B5. Amsterdam 2000. 209