Algoritmos para el pre procesamiento de nubes de puntos mediante representaciones dispersas

• Autores: Esmeide Alberto Leal Narváez
• Directores de la Tesis: John William Branch Bedoya (dir. tes.), Germán Sánchez Torres (dir. tes.)
• Lectura: En la Universidad Nacional de Colombia (UNAL) ( Colombia ) en 2020
• Idioma: español
• Enlaces
  • Tesis en acceso abierto en: repositorio.unal.edu.co
• Resumen

  • Nowadays, 3D scanners have become a standard source that provides millions of points as input for a growing number of applications areas such as industry, entertainment, medicine, computer vision, photogrammetry, etc. The large number of points generated by the scanning device is called point clouds. Point clouds often become complicated to handle due to multiple problems, such as the noise produced in the scanning process, lack of information (holes), or excess of information (points). Also, it is crucial in some applications to detect sharp features, like edges and valleys. We addressed all these problems in a stage of surface reconstruction called point cloud pre-processing.

    The focus of this thesis is the use of sparse representations for developing robust computational algorithms to solve the problems included in the pre-processing stage. Sparse representations are methods inspired in the human visual system, which can be adapted to the characteristics of problems found in the pre-processing of point clouds.

    We present contributions on some fundamental topics as denoising, sharp features extraction, hole detection, and simplification. With the direct use of the 3D points, we avoid the need for surface reconstruction methods, which are computationally complex and time-consuming.

    In this thesis, we introduce a smoothing method, which is effective in removing noise and preserving the sharp features and corners. The features preserving capability comes from the combining L1 median and L1 norm to estimate the normals and the point positions update.

    To reduce the sampling complexity of the cloud, we present a simplification method based on saliency. A Dictionary learning and sparse coding process are carried out over the normals and curvatures to find the saliencies. Next, it makes a selection of the sparse coefficients that represent the most salient features, carrying out in this way the simplification.

    We introduce a method for detecting features and holes in point clouds. First, we build a covariance matrix from the geometric information in a neighborhood around each point in the cloud. Then we estimate the eigenvalues of the covariance matrix, and combining them, we build feature vectors. The feature vectors are the signals to carry out a dictionary learning followed by a sparse coding process. At last, Imposing a threshold over the sparse coefficients, we detect features (edges, corners, valleys) and holes. We show the effectivity of our algorithms in a wide range of scanned geometric models of varying sizes, complexity, and details.