National Technical University of Athens (NTUA)
Δημιουργία φωτογραμμετρικού σαρωτή laser με χρήση πρόσθετων γεωμετρικών δεσμεύσεων
Abstract
dc:descriptionThis thesis presents the development, implementation and evaluation of a low cost surface scanner for small objects using additional constraints for increasing its robustness and accu¬racy. The scanner described here combines stereovision and the “slit-scanner” principle. Two web cameras, which are automatically pre-calibrated, view the trace of a moving hand-held laser plane in the scene. The laser line helps solve the correspondence problem and, finally, calcu¬late the coordinates of 3D surface points by using the simple parallax equations on epipolar images. Triangulation is then strengthened by extra geometric constraints. In more detail, the steps of the procedure are as follows: The system is calibrated automatically with bundle adjustment in order to estimate the in¬terior and the true-to-scale relative orientation of the two web cameras by processing stereo pairs of a 2D chess-board. For the purposes of this dissertation an algorithm was developed to automatically recognize the pattern coordinate system, collect all nodes of the chess-board and estimate initial values for all unknowns. This allows image pairs acquired during scanning to be re-sampled to (distortion-free) epipolar images. These images only depict the trace of the laser after subtraction of the background; also, smoothing filters are applied to reduce noise effects. The search for point homologies is thus confined on corresponding epi¬polar lines. Homologous points are defined as intersections of the epipolar lines with the viewed laser profiles (intensity peaks on each image row). Among a variety of tested methods, Gaussian curve fitting on each laser profile was chosen for peak detection. Peak position is given by si¬multaneously fitting three Gaussian profiles, with the additional curves being interpolated in the two diagonal directions through the initial peak approximation. This exploits informa¬tion from the peak neighbourhood, while still keeping the detected point on the epipolar row. The accuracy of peak position functions as criterion for point acceptance. The initial calculation of the 3D coordinates for all points by means of the parallax equations (solution with no redundancy) is strengthened by imposing extra geometric constraints to the solution: • All points reconstructed from the laser stripe of each image pair should belong to the laser plane. • Placing the object in a corner formed by two intersecting planes (which have been previously scanned) allows introducing an extra co-planarity constraint for all points lying on these background planes. • Neighbouring points should have compatible depth values (smoothing effect). The influence of these constraints, enforced either as ‘hard’ or ‘soft’ constraints, is examined and evaluated. The suggested solution estimates all unknown values of a whole scanning pro¬cess in a single adjustment step. The introduction of geometric constraints strengthens the solution by offering a redundancy of observations, thus increasing reliability, and also al¬lows addressing the problem of multiple intensity maxima due to occlusions or reflections. Fi¬nally, the point cloud generated can be augmented with RGB values by projection of back¬ground images generated automatically from the data set with a time median approach. In order to obtain a full 3D model of an object, several scans from different positions must be fused, using algorithms such as ICP. In the context of a fully automatic procedure, initial estimates for translation and rotation for ICP must also be automatically calculated. Here in¬formation form the background images is used, provided that the object is sufficiently tex¬tu¬red for an interest operator to extract sufficient points, both in number and in distribution. Using the SIFT and RANSAC algorithms, homologous points are extracted (via the fundamen¬tal matrix) between the images from adjacent scan positions. The 2D point pairs are upgrad¬ed to 3D by interpolation to the respective point clouds, thus making a rigid body transfor¬ma¬tion between the two point clouds possible. In the final step, the generated point cloud is smoothened. Several smoothing algorithms are evaluated and applied to the individual or the combined point clouds. For smoothing, it is as¬sumed that the object may be locally approximated by basic mathematical surfaces, while an anisotropic approach depending on local surface shape is also introduced. The accuracy of the scanner is in the order of 0.2 mm, as evaluated both theoretically and experimentally. For the practical evaluation, objects of known dimensions were scanned, whi¬le the system was also tested against a high-cost and a low-cost commercial scanner. The dissertation concludes with a presentation and evaluation of all applications carried out, its conclusions and suggestions, and finally the cited references.
Degree
thesis:*- Grantor dc:publisher
- National Technical University of Athens (NTUA)
- Year dc:date
- 2012
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Prokos, Antonios
- Πρόκος, Αντώνιος
Subjects
dc:subject × 22- Φωτογραμμετρία
- Σαρωτής τριγωνισμού
- Αυτοματοποίηση
- Βαθμονόμηση μηχανής λήψης
- Επιπολική γεωμετρία
- Γεωμετρικές δεσμεύσεις
- Συνόρθωση δέσμης
- Ανακατασκευή τρισδιάστατων επιφανειών
- Εξομάλυνση τρισδιάστατων επιφανειών
- Photogrammetry
- Slit scanner
- Automation
- Camera calibration
- Epipolar geometry
- Geometric constraints
- Bundle adjustment
- 3D surface reconstruction
- 3D surface smoothing
- Επιστήμες Μηχανικού και Τεχνολογία
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
- Engineering and Technology
- Electrical Engineering, Electronic Engineering, Information Engineering
Rights
- Language dc:language
- gre
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/26808