Einladung zum Statusvortrag im Promotionsver­fahren von Herrn Somnath Dutta

Abstract:

Accurate 3D point cloud registration is essential for robotics, computer vision, and graphics applications, where aligning partial shapes from different viewpoints is often necessary. Traditional methods, which rely on computationally expensive global alignment followed by local refinement, can be inefficient.

This work introduces rigid point cloud registration approaches using hierarchical surface descriptors based on arbitrarily fine sampling of continuous surface representations. To enhance performance, GPU-accelerated descriptor construction is introduced as an extension, enabling higher resolution and accuracy while reducing computational overhead. These methods minimize reliance on local refinement and are evaluated for their effectiveness in rigid registration tasks.

Additionally, the challenge of registering non-rigid, dynamic objects, particularly plants that undergo complex changes in topology and motion, is addressed using a tailored data association technique combined with a refined non-rigid registration algorithm. This method enables spatio-temporal analysis of plant dynamics, providing valuable insights into trait analysis.

Betreuer:        Prof. Dr. rer. nat. Stefan Gumhold
Fachreferent: Prof. Dr. Björn Andres