Abstract :
[en] The analysis of 3D changes in complex environments, such as urban and railway infrastructures, is essential for applications including digital twins, infrastructure monitoring, and land management. While multi-temporal point clouds provide rich spatial and temporal information, their unstructured nature, lack of semantics, and acquisition inconsistencies make automated change detection a challenging task. This thesis addresses these challenges by proposing a comprehensive object-based framework that integrates semantic segmentation, change indicators, and structured modeling to improve the reliability, interpretability, and interoperability of 3D change detection.
The thesis reviews existing 3D change detection methods, highlighting key challenges related to noise sensitivity, lack of semantic interpretation, and limited integration into structured outputs. In response, a semantic and geometric fusion method is introduced. Bi-temporal point clouds are first semantically segmented, followed by Cut-Pursuit clustering to extract spatially coherent object instances. Second, object-level change indicators, including geometric descriptors such as verticality, sphericity, and omnivariance, are computed and used to classify changes into four categories: appeared, disappeared, modified, and unchanged. Third, a Random Forest classifier is trained to predict change labels, demonstrating improved spatial coherence and robustness to noise compared to point-wise approaches.
To support practical applications in city modeling and infrastructure documentation, the thesis introduces a rule-based workflow that formalizes detected changes into a CityJSON representation. This structured output encodes temporal, geometric, and semantic attributes of changed objects without requiring labeled training data. The approach allows for efficient updates of digital twins and facilitates integration into 3D geographic information systems. The framework is also extended to the monitoring of railway environments. A novel annotated dataset, Rail3D, is created to support semantic segmentation of railway-specific classes. Deep learning models are benchmarked on this dataset, and the results are used to detect and analyze vegetation changes over time. This component highlights the applicability of the proposed methods for infrastructure maintenance and monitoring tasks.
This thesis demonstrates how semantic enrichment, geometric analysis, and structured representation can be effectively combined to advance the state of 3D change detection. The developed methods are validated on real-world applications and contribute to the automation and interpretability of 3D point cloud analysis in dynamic environments.
