Complex scenes can be digitized in 3D point clouds using LIDAR devices. These scenes include recurrent objects that can be segmented and grouped by class. Those objects can be static, dynamic or changing periodically in the scene.
However the clouds produced by the scanner are challenging: they are unstructured and can present perturbation such as varying densities, noises and partial occlusion on some parts.
During this thesis, we will address the problem of point cloud semantic segmentation using neural network approaches with specific inputs datas, such as the point cloud, but also the information from the pictures taken by the scanning device.
This project is part of the industrial development from the Technodigit company on the production of digitized large complex scenes, in partnership with the Origami team from the LIRIS.