Actors: Inria / TITANE, Cerema / ENDSUM, Cerema / DTerSO, Cerema / GIPI, Inria / STATIFY, Cerema / GITEX, Cerema / DTerMed, Cerema / DTerCE, Cerema / GéoCoD.
3D modeling concerns structures of very diverse interest such as structures (bridges, tunnels, walls, dikes), pavements, soils, cliffs … Hence their protean character. A first problem is the exhaustive and automatic census of works at the scale of a territory. The objective being to understand the evolution of structures and the environment such as ground movements, a spatio-temporal analysis is required, with the scientific challenge of very slow movements, of the order of measurement noise. The metrological analyzes require distance calculations, the latter requiring the readjustment of the acquisitions on the same site, as well as the matching of the acquisitions made at different times. Finally, the non-destructive analysis of structures, pavements and materials requires modeling and reasoning with 3D volumes.
Task 2.1: Detection of structures from cartographic data
Actors: Inria / TITANE, Cerema / ENDSUM, Cerema / DTerSO, Cerema / GIPI.
The objective is to automatically detect the probable position of structures from satellite or aerial images, at the scale of a territory, in order to plan field visits. The general hypothesis that a structure makes it possible to overcome an obstacle on a road, rail or river communication route, suggests exploring the detection of lane intersections. However, several pitfalls make the problem difficult: intersections are not always visible and the level differences with the lanes are sometimes very small. As the presence of a book depends on a context that is not necessarily local, we will explore supervised learning methods.
Task 2.2: Registration and semantic segmentation of 3D data
Actors: Inria / TITANE, Inria / STATIFY, Cerema / ENDSUM, Cerema / GITEX, Cerema / DTerMed, Cerema / DTerCE, Cerema / GéoCoD.
The 3D data capture carried out in axis 1 will generate massive data such as unstructured 3D point clouds, 3D point lines or even images. Several acquisitions are necessary to cover an entire structure or environment, and spatio-temporal monitoring requires making acquisitions at different time intervals. For a given site, the registration and matching of this data is an essential step before the metrological analysis and in particular the calculation of distances. The registration of 3D point clouds consists in estimating a rigid transformation which aligns all the clouds in the same frame of reference to form a single cloud that can be analyzed. The scientific challenge is to perform the estimate in the presence of varying resolutions, measurement uncertainties (variable noise, weather occultations, outliers), and significant variations in unstable environments. The matching of point clouds or surfaces acquired at different times consists in calculating a univocal (ideally bijective) function between each point of a cloud – or of a surface reconstructed from the cloud – and its analogue on the other clouds. The difficulty is to be robust (stable) in the presence of missing or aberrant data, while allowing the calculation of distances and associated reliability. We will explore diffuse rather than point-to-point correspondence functions, and robust distances based on optimal transport plans relaxing the constraints of mass preservation. A dense semantic segmentation method will also be designed to facilitate registration and matching, including for the geographic registration of measurement data, that is to say their positioning in the global frame of a vector geographic map or halftone. Semantic segmentation and the use of other types of measures (such as color or reflectivity) make it possible to consider methods of exhaustive exploration of the transformation space (so-called global registration methods) rather than local searches for correspondence. This exploration will make it possible to obtain an alignment between clouds of points which are not pre-aligned.
Task 2.3: 3D internal mapping of soils and pavements
Actors: Inria / TITANE, Inria / STATIFY, Cerema / ENDSUM, Cerema / DGIPI.
Most of the current methods work by analyzing images, line or area data. The problems of soil mapping and non-destructive analysis of structures, pavements (including thicknesses, internal cracks, detachments and buried networks) and materials require the acquisition and analysis of 3D volume data, by tomography or radar imagery or electrical resistivity. The volume analysis of this data requires processing and modeling volumes and reasoning in 3D for the analysis. The analysis includes dense segmentation and the detection of defects such as internal cracks and spatiotemporal deformations. The modeling includes the generation of multi-domain meshes to facilitate modeling and simulation of the rock hazard for prediction from map and meteorological data.
Task 2.4: 3D internal mapping of structures
Actors: Inria / TITANE, Inria / STATIFY, Cerema / ENDSUM, Cerema / DGIPI, Cerema / GITEX.
As for pavements, the mapping problems of civil engineering structures require the acquisition and analysis of 3D volume data, by tomography or radar, ultrasound or electrical resistivity imaging (for example to improve knowledge of the internal geometry of the structures. structures, to address issues of material durability, for the auscultation of inaccessible areas and the detection of hidden defects, etc.). The analysis includes dense segmentation and the detection of defects such as internal cracks and spatiotemporal deformations. The modeling includes the generation of multi-domain meshes to facilitate the modeling and simulation of quantitative quantities (water and chloride content gradients, porosity, characterization of the state of internal stresses, risk of sudden rupture, etc.).
