Object identication in 3D urban environments

Abstract: This project investigates the problem of object identification in three-dimensional (3D) urban models represented by meshes. More specifically, the objective is to detect defects or poorly rendered objects, called artefacts, in order to remove them later on. The literature on related urban analyses is marginal for meshes while abundant for 3D point clouds. The contribution of this project is then twofold: studying whether objects can be identified in meshes and how semantic mesh segmentation methods can be extended to lower-resolution meshes. Our methods correspond to alternative solutions for different steps of an unsupervised pipeline algorithm commonly used for object classification in 3D point clouds. First, a ground model is generated either from an elevation image-based approach or from direct clustering of the triangles. The latter corresponds to a mesh segmentation problem and was investigated using either k-means or a Markov Random Field formulation. The clustering approach divides the input mesh in different meshes with the following classes: ground, façade, roof and optionally vegetation. The project investigates two new features that can help identify vegetation in lower-resolution meshes. Then, objects are segmented from the ground model using a watershed approach with local maxima as markers and additional propagation constraints based on textures. Our results include a survey study based on users’ visual inspection in order to evaluate methods against one another. Our findings can be put in the context of urban meshes’ semantic analysis. Artefacts in urban meshes are successfully detected by extending already existing mesh segmentation methods in association with a density-based feature. The survey results also support the hypothesis that existing mesh segmentation methods do not adapt well to lower-resolution meshes. Finally, regardless of mesh resolution, successful vegetation identification is the main remaining problem in most approaches.