Uma investigação sobre o uso de descritores geométricos de superfície para filtragem de regiões visando o aumento da qualidade do registro de nuvens de pontos 3D.

• Ano de defesa: 2023
• Autor(a) principal: Souza, Daniel Araújo Chaves
• Orientador(a): Soares, José Marques
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Área do conhecimento CNPq: CNPQ::ENGENHARIAS::ENGENHARIA ELETRICA
• Link de acesso: http://repositorio.ufc.br/handle/riufc/74689

With a broad application in various domains such as robotics, computer vision, and reconstruction, the registration of 3D point clouds presents challenges in terms of computational efficiency and the quality of alignment obtained, especially in point clouds with a high number of points and partial views captured by sensors from different perspectives. In this context, various approaches suggest optimizing registration by selecting relevant subsets of points from the original point clouds. However, determining the most suitable regions for registration is not a straightforward task. In this regard, an investigation is proposed regarding the most suitable regions for the registration of point clouds using geometric descriptors derived from the covariance matrices of point neighborhoods. The goal is to identify more effective selection strategies to enhance alignment in various scenarios. This study comprehensively and in detail documents the application of surface descriptors in the selection of relevant regions for registration, along with effective strategies for incorporating these filters into a registration pipeline. The performance evaluation of the investigated strategies was conducted by calculating the transformation error, obtained through ground truth, ensuring a precise analysis of the quality of the achieved alignment. The results of various experiments conducted, spanning from object models to indoor and outdoor scenes, showed that the selection of regions of interest based on surface descriptors results in a lower average transformation error. This leads to a greater number of samples satisfactorily aligned according to the quality criteria established in this study, compared to the evaluated classical registration strategies, especially in the context of object models.