Segmentation sémantique de nuages de points multisources de parcelles et d’ouvrages existants à l'aide de techniques d'apprentissage machine

Abstract

Remote acquisition technologies are becoming more and more accessible and are no longer reserved for professionals. They are therefore being used more frequently and opportunities to use them are only increasing. These technologies, coupled with the growing computing power of computers, make it possible to acquire and process a large amount of data. Combining data from different technologies is therefore becoming possible. This research project aligns with this topic and involves the digitization of large geographic areas ranging in size from several hundred square meters to one square kilometer with several acquisition technologies. In our case, we had access to surveys from four distinct acquisition sources: airborne lidar onboard aircraft and helicopters, photogrammetry from UAVs and terrestrial lidar. The main requirements to fulfill the project are to be able to create a complete point cloud from several types of sources and be able to semantically segment multisources point clouds. These requirements raise the question of how best to use the various sources available to create and segment a multisource point cloud. The use of multiple data sources that provide heterogeneous data for the same project is not well covered in the literature for a variety of reasons. One of the main reasons is that there are very few projects involving multiple technologies in the industry. The acquisition technologies that are selected for a project are those that best meets the project’s needs. The idea of using only one technology per project is increasingly being challenged with the use of UAVs and photogrammetry to support ground-based lidar. However, the body of scientific literature currently includes few articles that take into account multiple sources and even fewer that use other combinations of sources than photography + lidar. This project therefore provides multiple contributions. First, it includes multiple modules that have been developed to process and semantically segment multisources point clouds. Second, the propagation of the attributes acquired by the different sources on the neighboring points and, finally, a simple system to take into account the neighborhood. Our proposed method enabled us to semantically segment multisources point clouds by taking into account the attributes of each source using different modules. An ablation study was then performed to evaluate the interest of the different modules in our proposed method. Our method provided very good semantic segmentation results in comparison with algorithms used in existing works on this subject. Our work aligns with the openness to complementary sources that is observed in digital capture projects.

Date	27 Jul 2023
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Louis Rivest (Supervisor), Antoine Tahan (Co-supervisor), Jean-Philippe Pernot (Co-supervisor) & Arnaud Polette (Co-supervisor)