Carregant...
Tipus de document
TesiVersió
Versió publicadaData de publicació
Si us plau utilitzeu sempre aquest identificador per citar o enllaçar aquest document: https://hdl.handle.net/2445/217096
Advancing 3D Point Cloud Understanding in Real-World Applications
Títol de la revista
Autors
ISSN de la revista
Títol del volum
Resum
[eng] In recent years, rapid advancements in 3D sensing technologies, such as LiDAR, have revolutionized various fields, spanning robotics, augmented reality, earth vision, and industrial manufacturing. Point clouds, acquired through 3D sensing technologies, capture the geometric structure of objects and environments in a three-dimensional coordinate system. However, the lack of structure and order in point clouds, as well as the significantly redundant and inconsistent sample densities, makes it challenging to analyze. In addition, point clouds captured in real-world setups contain millions of points accompanied by noise either from the measurement tool itself or by moving objects in the scene, leading to further computational challenges. Moreover, point clouds are frequently obtained dynamically, generating temporal 3D point clouds, where the alignment between frames is often essential for spatial consistency and accurate analysis. Consequently, learning intelligent models on either static or temporal 3D point clouds presents important challenges and necessitates advanced computational techniques to address issues of accuracy, efficiency, and robustness depending on the application needs. This thesis presents research efforts toward advancing 3D point cloud understanding, providing theoretical and practical insights to enhance decision-making and application in fields where spatial and geometric understanding is important. The main contributions of this thesis are seven, divided into three thematic parts, Model Evaluation and Interactive Visualization in 3D Point Cloud Segmentation, 3D Understanding Applications in Geoscience and 3D Understanding Applications in Industrial Manufacturing. Concerning the first thematic part, four contributions have been made. The first contribution includes a benchmarking of point cloud segmentation models that instead of focusing solely on accuracy-related performance metrics, further incorporates time and memory efficiency evaluation. Then, our second contribution encompasses a comprehensive analysis of state-of-the-art deep learning architectures for 3D point cloud segmentation. Through rigorous evaluation approaches, novel performance metrics are proposed to facilitate effective model comparison, considering ac- curacy, time and memory efficiency, and robustness. In addition, we create a correspondence between model design properties and experimental properties, further elucidating model selection. Our third contribution includes a proposed early-stopping technique for enhancing the trade-off between efficiency and accuracy in training neural networks on point cloud segmentation. Our fourth contribution to model evaluation and selection is visual insights into the results of point cloud segmentation models during and after the learning process. A new dashboard visualization tool, named CLOSED, has been proposed facilitating the rigorous comparison of different neural networks in 3D point cloud segmentation. Concerning the second part on applications in the intersection of Geoscience, Machine Learning, and Computer Vision, two contributions have been made towards automating rockfall detection utilizing real-world captured temporal 3D point clouds. The first one focuses on enhancing rock- fall detection and addresses critical challenges such as class imbalance in the detection of rockfall candidate areas. Initially, various machine learning algorithms are studied alongside resampling techniques on real-world 3D scans in order to develop an intelligent framework for rockfall detection. This framework, further extended to incorporate geological properties, demonstrates high accuracy and robustness in detecting areas susceptible to rockfall. The second contribution leverages advancements in point-based neural networks and spatiotemporal information from the captured 3D scans to improve the accuracy and efficiency of rockfall candidate area detection. The proposed method showcases effectiveness in identifying rockfall candidate areas directly from real-world 3D scans. Finally, concerning the third part on applications in the intersection of Industrial Manufacturing, Deep Learning, and Computer Vision, one contribution has been made. Our contribution lies in adapting 3D point cloud understanding models to accurately identify various machining tools based on sampled surfaces, enabling insights for optimizing industrial machining processes and reverse engineering workpieces. Specifically, we propose a novel process and develop guidelines for the task of machining tool identification using temporal 3D point clouds, sampled from the tool engagement surface from a workpiece in progress. This thesis advances 3D point cloud understanding in real-world applications. It addresses existing gaps in the research field, poses new research questions, and explores novel research directions. We present deep learning and classical machine learning techniques on 3D point clouds and real-world applications in geoscience and industrial manufacturing. The outlined contributions establish a basis for further advancements and effective utilization of models on 3D point clouds across diverse disciplines. Our findings, developed software, and resources presented in this dissertation are available to the community to facilitate further research and knowledge transfer.
Descripció
Matèries (anglès)
Citació
Citació
ZOUMPEKAS, Athanasios. Advancing 3D Point Cloud Understanding in Real-World Applications. [consulta: 30 de novembre de 2025]. [Disponible a: https://hdl.handle.net/2445/217096]