TUD Logo

TUD Home » ... » Results of student work » Minor Theses » André Jähnig

Computer Graphics

Minor thesis of André Jähnig

Matching and registrierung of 3D-scans

Lehrstuhl für Computergraphik und Visualisierung

Student: André Jähnig
Tutor: Dipl.-Medien-Inf. Sören König
Professor: Prof. Dr. Sefan Gumhold


In recent years claims on digital 3d-models increased more and more and due to better computer-performances more realistic models can be created. In addition to complex modelling of such objects, you can also scan the objects with a 3d-scanner and reconstruct them with the help of a computer. To enable this process a software was created for the high-speed-scanner, developed by the professorship. This software aligns single scans, which represent only partial sections, one to another, resulting in a three-dimensional pointcloud, which describes the whole object scanned. the problem of this process: the scanner does not provide any information about the movements of the object or the scanner.


As the main basis for solving this problem the standard solution was chosen: the iterative closest point algorithm (ICP) by Besl and McKay. It aligns two scans respectively two pointclouds one after the other. With the help of the euclidean distance the algorithm searches for corresponding pairs of points and determines a transformation to bring them together spatially. But there is a big handicap: the algorithm does not function if the two scans only overlap partially. That's why different modifications and additional features were implemented so that the alorithm got faster and more stable. Besides this process, a global registration-method was implemented which uses not only two but all scans at once, so that all scans will be aligned parallelly.


Because we have to deal with data of a high-speed-scanner working currently with 17fps, it doesn't make any sense to use all scans for the registration. That's why only those scans are selected that vary to a certain level. You can see the registrations of two different figures below, with scans consisting of 200.000 - 300.000 points each. The bust was created of 14 scans and its calculation took about 3 minutes. Despite the ICP-algorithms' modifications, the biggest problems have always been outliers. Outliers are parts of the scan and a result of an incorrect acquisition. But they do not represent any surface of the object.

Büste 01 Büste 02
Büste 03 Büste 04
Ergebnis der Registrierung von 14 verschiedenen Scans einer Büste
Frosch 01 Frosch 02
Frosch 03 Frosch 04
Ergebnis der Registrierung von 20 verschiedenen Scans einer Frosches

Future work

The presented procedure could be improved, e.g. by a detection and elemination of outliers previous to the registration. There is also the possibility to analyze how to improve the speed of the single steps of the calculation. An attempt could be the caching of detected corresponding points during the creation of the point-pairs. Furthermore the use of other registration-methods, for example those working with range-images, should be taken into consideration. After the registration the surface-reconstruction would follow. Thereby the scanned object could be displayed with the help of a polygonal mesh or an volumetric representation.


Last modified: 1st Feb 2010, 2.38 PM
Author: Dipl.-Medieninf. Sören König