Towards a mesh based, structured 3D representation of cities

Overview & Synopsis

This thesis focuses on the 3D reconstruction of urban environments using exclusively textured triangular meshes. This mesh format was chosen for its universality, as it has become a standard output of most photogrammetric processing chains. These meshes effectively combine both geometric and radiometric data to support three-dimensional geographic information in a compact and easy-to-use format. The main objective of this research work is to transform these textured meshes into structured 3D models. We introduce the concept of structured mesh model. In a structured mesh, each face of the mesh is associated with metadata such as its semantic class (ground, vegetation, rail, road, sidewalk, etc.) and instance information. Each object then consists of a subset of mesh faces of the same instance. This approach allows for the faithful 3D visualization of the world and paves the way for numerous applications in simulation and urban planning.The first major contribution of this thesis is the study and development of methods for the semantic segmentation of these textured meshes. This segmentation is essential for the correct assignment of categories to the different faces of the mesh, forming a fundamental basis for subsequent stages of our work.The second main contribution is a method for simplifying segmented meshes. This simplification aims to reduce the complexity of the mesh while preserving geometric and semantic information, enforcing the alignment of mesh edges with object boundaries and semantic class divisions. The aim is to clearly identify each object and associate it with the simplest possible geometry, so as to obtain a model with a good compromise between accuracy and complexity, thus saving storage space and calculation time for subsequent visualization or simulation algorithms. The final result of this process is a ready-to-use structured mesh. The third significant contribution of this thesis is the augmentation of a structured mesh's fidelity in occluded areas, such as under bridges, tunnels, and dense vegetation. By focusing particularly on networks like roads, rails, and rivers, we leverage their intrinsic characteristics to improve the precision and reliability of our structured meshes in these complex areas, with the final goal of making these networks navigable. This thesis therefore proposes a new approach to city modeling, in the form of a complete pipeline for automatically generating a structured model containing navigable networks from a textured mesh