Eingang zum Volltext in OPUS
Hinweis zum Urheberrecht
Dissertation zugänglich unter
3D Global Shape Descriptors Applied in Scan Registration
3D Global Shape Deskriptoren in Scan Eintragung beantragt
Dokument 1.pdf (4.120 KB)
Freie Schlagwörter (Englisch):
3D data , shape descriptor , scan registration , mobile robotic mapping
Zhang, Jianwei (Prof. Dr.)
Tag der mündlichen Prüfung:
Kurzfassung auf Englisch:
With the advent of low-cost and high-performance depth sensors, the usage of three-dimensional (3D) point clouds is becoming attractive and increasingly popular. Meanwhile, there is a growing demand to process and understand the 3D data. 3D scan registration is the cornerstone of several advanced 3D data processing techniques, and the odometry-free scan registration of 3D point clouds is a research hot spot recently. For the present, the majority of state-of-the-art registration methods depend on reliable initial estimates or local salient features. The 3D registration techniques based on global shape descriptors have not attracted as much attention as they deserve. Generally speaking, the original 3D data is depicted by a global descriptor, and the transformation between the original data could be solved by aligning the corresponding global descriptors. In this thesis, we propose two novel registration methods based on two global shape descriptors, namely Hough Transform Descriptor and Spherical Entropy Image respectively.
For the Hough Transform Descriptor-based registration method, the original 3D scans are projected into the Hough domain. In this way, 3D rotation of the original data is decoupled from its 3D translation, and then the rotation and translation between the original data could be recovered separately. The rotation is estimated firstly by aligning the corresponding translation invariant Hough Transform Descriptors, and there is only translation between the original data after rotating them according to the estimated rotation matrix. The Phase Only Matched Filter (POMF) is employed to estimate the translation between the rerotated scans. We also propose a novel shape descriptor named Spherical Entropy Image (SEI), and develop a novel registration method based on SEI aided by the Spherical Harmonic analysis techniques. Since SEI is not translation-invariant, it is impossible to estimate the rotation and translation separately as the aforementioned Hough Transform Descriptor-based registration method does. In our SEI-based registration algorithm, we integrate the rotation estimation and translation recovery into an iteration framework, which is one of our major contributions. Besides, the possibility of using SEI as a local shape descriptor in feature matching task is also discussed.
Elaborate experiments with regard to public available datasets and the dataset captured by our custom-built platform are implemented to validate the efficiency of our proposed registration algorithms. The experiment results illustrate the parameter-insensitivity, runtime stability, high reliability and efficiency of our novel algorithms in the registration of feature-less, partially overlapping and largely transformed 3D scan pairs.