Mainz University of Applied Sciences, as the coordinating university, is conducting a research college in the competence area of artificial intelligence together with the University of Koblenz over a period of three years. With this college, financed by the funding line Forschungskolleg Rheinland-Pfalz, the already existing cooperation with the University of Koblenz in the research field of artificial intelligence is to be bundled and expanded.
The six professors involved belong to the Faculty of Computer Science of the University of Koblenz, as well as to the Departments of Engineering and Economics of the Mainz University of Applied Sciences. A total of five doctorates are to be produced within the framework of the college. One of the doctoral theses, entitled Adaptive registration and semantic interpretation of point clouds in indoor environments, will be produced at i3mainz under the supervision of Thomas Klauer.
The rapid development of low-cost sensors and 3D data acquisition technologies has increased the importance of point cloud analysis and interpretation in various application areas such as robotics, virtual reality, and architectural design.
Interest in indoor applications including smart buildings (BIM), indoor navigation, and asset management is growing, highlighting the need for accurate modeling and understanding of the indoor environment.
In the context of smart buildings (BIM), indoor navigation, and asset management, point clouds serve as an essential resource for capturing spatial information. Nevertheless, point clouds in indoor environments pose a particular challenge due to occlusions, scene dynamics, and geometric variations.
The research work titled Adaptive Registration and Semantic Interpretation of Indoor Point Clouds attempts to develop a hybrid framework for adaptive registration and semantic interpretation of indoor point clouds collected by low-cost sensors such as iPhones.
Current approaches to indoor point cloud registration and semantic interpretation largely focus on deep learning or knowledge-based methods. However, these approaches face limitations when processing noisy data with geometric variations and incomplete data. In particular, deep learning approaches require extensively annotated datasets for training, which is very challenging, especially when dealing with rare use cases or new sensors. Furthermore, knowledge-based methods often require significant input from domain experts to adequately describe knowledge.
The PhD aims to overcome the limitations of existing approaches by combining deep learning techniques with knowledge-based methods. To integrate both approaches, distributed high-dimensional vector representations can be used as a unifying representation substrate. The proposed framework will incorporate technologies such as NERF (Neural Radiance Files) and self-learning ontologies to extend and, where appropriate, improve the semantic interpretation of indoor point clouds.
By addressing challenges associated with registering multi-part point clouds and understanding dynamic scenes, this research will contribute to the development of more robust and accurate methods for indoor understanding.