2026-01-292026-01-292025-04-04RIBEIRO, Thiago Figueiró. 3D geometric reconstruction of civil infrastructures with neural radiance fields. Orientador: João Crisóstomo Weyl Albuquerque Costa. 2025. 96 f. Dissertação (Mestrado em Engenharia Elétrica) - Instituto de Tecnologia, Universidade Federal do Pará, Belém, ano de defesa. Disponível em:https://repositorio.ufpa.br/handle/2011/17905 . Acesso em:.https://repositorio.ufpa.br/handle/2011/17905The use of noncontact sensing technologies for structural health monitoring (SHM) has signifi- cantly broadened the scope of tools available for precise measurement and analysis in engineering and scientific contexts. They address several limitations of the conventional contact-based sen- sors and at times outperforming them while easier survey, more convenient to install and often low-cost. Digital Twins — dynamic, data-driven virtual replicas of physical structures — have further revolutionized SHM by integrating real-time sensor data with predictive analytics and computational modeling. LiDAR and photogrammetry technology are leveraged to build high- fidelity 3D reconstruction models which can be used to create building information models and digital twins for civil structures. Recent deep learning advancements marked a paradigm shift in several areas, including 3D reconstruction. One particular approach is using Neural Radiance Fields, a deep-learning-based methodology capable of producing high-fidelity 3D models from sparse image datasets, such as those captured using standard consumer-grade cameras or smartphones. NeRF is capable of generating dense point clouds comparable to those generated by Multiview Stereo photogrammetry and terrestrial laser scanning. However, there is a gap in literature addressing the quantitative capabilities of NeRF-based 3D scanning of bridges. This work evaluates the performance of NeRF 3D reconstructions of real-world bridges against SFM/MVS photogrammetry-based models and ground truth data generated via LiDAR. We demonstrate NeRF’s feasibility for large-scale structural assessments, with key insights into its performance under varying data availability, impacting both LOA (Level of Accuracy) metrics and error measurements. NeRF offers higher LOA and lower standard error, mean average error and when compared to Photogrammetry. Overall, NeRF proves to be the more robust and accurate method, especially when balancing data availability with reconstruction quality, positioning it as the preferred choice for 3D modeling in data-limited, budget-constrained, specialized-equipment-scarce scenarios. This approach offers an efficient, cost-effective, and accurate solution to address the increasing demands for infrastructure monitoring within the Architecture, Engineering and Construction industry, thereby reducing economic and technical barriers to adoption.enAttribution-NonCommercial-NoDerivatives 4.0 Internationalhttp://creativecommons.org/licenses/by-nc-nd/4.0/Neural radiance fieldsEscaneamento 3DNúvens de ponto 3DRenderização volumétricaSHMNeural radiance fields3D scanning3D point cloudsVolumetric renderingBuilding information modelingDissertaçãoCNPQ::ENGENHARIAS::ENGENHARIA ELETRICAINTELIGÊNCIA COMPUTACIONALCOMPUTAÇÃO APLICADA