Object recognition in road and urban environments from massive datasets collected by mobile mapping sensors

Laburpena

During the last decade there has been a significant improvement in infrastructure surveying technologies and mobile mapping systems. It is now possible to acquire dense and tridimensional representations of a surveyed environment which capture geometric and radiometric data in the form of unorganized sets of 3D points called point clouds. The conversion of these unorganized data into meaningful information is not straightforward for the user, and requires the development of specific algorithms for solving specific problems. In that context, this thesis proposes a set of methodologies which are mainly aimed at assisting road network inventory and maintenance processes, automatizing them and increasing their objectivity by avoiding the unique application of maintenance operators' criteria. Specifically, the thesis focuses on two of the most important road network assets which are vertical traffic signs and road markings. In essence, the proposed methodologies have to be able to detect and classify these elements from data acquired by an aerial or terrestrial mobile mapping system, which mainly consist of 3D point clouds and 2D imagery. Several challenges have to be tackled such as data organization and storage, development of preprocessing frameworks, point cloud structure understanding and organization, generation of machine learning models, or output visualization and exportation. Ideally, all the meaningful information about road network assets which includes, for instance, geographic position, orientation or meaning of surveyed traffic signs, will update an inventory database and optimize the maintenance activities on the road. Every work in this thesis has been tested within carefully selected study cases, obtaining state-of-the-art results and therefore contributing to the knowledge in its respective scope. This thesis is structured as a compendium of seven scientific publications, all of them published in high impact, peer reviewed, international journals indexed on the Journal Citation Report (JCR).