Thesis defense by Jérémy Montlahuc, doctoral student at the LISPEN : "Semantic segmentation of multi-source point clouds of existing plots and structures using machine learning techniques."
Summary
Remote acquisition technologies are becoming increasingly available and powerful. Their deployment as contactless measurement technology is becoming more common, and the possibilities for their use are growing. These technologies, coupled with the computing power of computers, make it possible to acquire and process large amounts of data. Combining data from various technologies is therefore becoming feasible, and even desirable, in order to overcome the limitations of each technology.
This research project is part of this trend. Its field of application is the digitization of large geographical areas, ranging from 10² m² to 10 km², using several acquisition technologies simultaneously. In our specific case, we have access to data from four distinct acquisition technologies: airborne lidar on board aircraft, lidar on board helicopters, drone photogrammetry, and terrestrial lidar. The main requirement is the automated creation, from several types of surveys and via a semantic segmentation algorithm, of a complete and consolidated multi-source point cloud. This raises the fundamental question of how to best exploit the different surveys, which vary in density and characteristics, in order to create and accurately segment the multi-source point cloud.
The various data sources provide heterogeneous data. Their fusion and consolidation are topics that are rarely addressed in the literature. The main reason for this is the lack of industrial projects involving multiple acquisition technologies. When these technologies are implemented simultaneously for a project, they must meet criteria that can be summarized as a coherent cloud and precise segmentation. The use of a single technology per project is increasingly being called into question. Indeed, with the growing use of drones and aerial photogrammetry to support terrestrial lidars, there is increasing access to so-called multi-source data capture. However, the scientific literature is currently lacking in articles that consider multiple sources, and is even more scarce when it comes to sources other than a combination of photography and lidar.
Our work is part of an effort to open up digital capture projects for use with various complementary sources. Our main contribution is a set of modules developed to process and semantically segment multi-source point clouds. The propagation of attributes acquired by the different sources is based on the notion of proximity. Neighboring points are interlinked and attributes are thus consolidated, which has made it possible to segment semantically while benefiting from the advantages of each technology. An ablation study is carried out to evaluate the value of the various modules in our proposal. Our proposal is validated on multi-source point clouds provided by our industrial partner. In addition, the performance of our proposal is compared favorably with other algorithms recently proposed in the scientific literature on this subject. Our proposal has achieved better semantic segmentation performance. However, the computation time is slightly longer.
Composition of the jury
Mr. Louis RIVEST, thesis supervisor, Professor, Department of Systems Engineering at the École de technologie supérieure
Mr. Antoine TAHAN, thesis co-supervisor, Professor, Department of Mechanical Engineering at the École de technologie supérieure
Mr. Jean-Philippe PERNOT, co-supervisor of thesis, University Professor, LISPEN, Arts et Métiers University
Mr. Arnaud POLETTE, co-supervisor of thesis, Associate Professor, LISPEN, Arts et Métiers University
Mr. Jean-Luc MARI, rapporteur, University Professor, LIS, Faculty of Science, Aix-Marseille University
Mr. Yann QUINSAT, rapporteur, Associate Professor, LURPA, at Paris-Saclay University
Mr. Marc-Antoine DROUIN, External Examiner, NRC
Mr. Michel GUÉVREMONT, external examiner, Hydro-Québec
The defense will take place on Tuesday, June 27, 2023, at 3:00 p.m. via Zoom.
