Abstract
The development of MLS LiDAR acquisition with devices mounted on vehicles or drones, make it possible to scan entire cities in the form of georeferenced three-dimensional point clouds. The use of these data by city managers allows the census as well as the monitoring over time of urban objects whether they are fixed (lampposts, bus shelters, etc.), mobile (garbage containers) or natural (trees) in order to be able to intervene in the event of disappearance, displacement, deterioration or potential danger.This approach requires being able to process large clouds that can have several hundred million points and bring together thousands of objects. It therefore becomes necessary to automate the processing applied to point clouds in order to be able to automatically extract and classify the elements which correspond to urban objects. The diversity as well as the large number of urban objects present in cities are a real challenge for the development of automated approaches. In this thesis we explore the recent path of deep learning applied to unstructured data to achieve the localization and automatic classification of urban objects in a 3D point cloud. Drawing inspiration from recent advances made possible by the PointNet network, we propose to carry out supervised learning directly from the raw point clouds without applying intermediate transformations. We have thus developed a 3D neural architecture that we have based on an original layer allowing simultaneously grouping points and extracting characteristics from them. From this architecture, we present the results we obtained on the task of detecting urban objects in LiDAR point clouds scanned in the streets of large cities.