The contribution of Deep Learning to the semantic segmentation of 3D point-clouds in urban areas

Abstract

Semantic segmentation in a large-scale urban environment is crucial for a deep and rigorous understanding of urban environments. The development of Lidar tools in terms of resolution and precision offers a good opportunity to satisfy the need of developing 3D city models. In this context, deep learning revolutionizes the field of computer vision and demonstrates a good performance in semantic segmentation. To achieve this objective, we propose to design a scientific methodology involving a method of deep learning by integrating several data sources (Lidar data, aerial images, etc) to recognize objects semantically and automatically. We aim at extracting automatically the maximum amount of semantic information in a urban environment with a high accuracy and performance.