Novas abordagens para segmentação de nuvens de pontos aplicadas à robótica autônoma e reconstrução 3D

Nenhuma Miniatura disponível
Título da Revista
ISSN da Revista
Título de Volume
Universidade Federal de Goiás
Depth sensing methods yield point clouds that represent neighboring surfaces. Interpreting and extracting information from point clouds is an established field, full of yet unsolved challenges. Classic image processing algorithms are not applicable or must be adapted because the organized structure of 2D images is not available. This work presents three contribution to the field of point cloud processing and segmentation. These contributions are the results of investigations carried out at the Laboratory for Education and Innovation in Automation – LEIA, aiming to advance the knowledges related to applying spacial sensing to autonomous robotics. The first contribution consists of a new algorithm, based on evolutionary methods, for extracting planes from point clouds. Based on the method proposed by Bazargani, Mateus e Loja (2015), this contribution consists of adopting evolutionary strategies in place of genetic algorithms making the process less sensitive to user-defined parameters. The second contribution is a method for segmenting ground and obstacles from point clouds for autonomous navigation, that utilizes the proposed plane extraction algorithm. The use of a quadtree for adaptive area segmentation allows for classifying points with high accuracy efficiently and with a time performance compatible with low cost embedded devices. The third contribution is a variant of the proposed segmentation method that is more noise tolerant and robust by incorporating a neural classifier. The use of a neural classifier in place of simple thresholding makes the process less sensitive to point cloud noise and faults, making it specially interesting for processing point clouds obtained from real time stereo reconstruction methods. A through sensitivity, accuracy, and efficiency analysis is presented for each algorithm. The dihedral angle metric (angle between the detected plane and the reference polygons that share at least one point) proposed by Bazargani, Mateus e Loja (2015) is used to quantify the plane detection method accuracy. The ratio between the correctly classified points and the total number of points is utilized as an accuracy metric for the ground segmentation methods. Additionally, computing costs and execution times are considered and compared to the main state-of-the-art methods.
MARCON S., G. A. Novas abordagens para segmentação de nuvens de pontos aplicadas à robótica autônoma e reconstrução 3D. 2016. 112 f. Dissertação (Mestrado em Engenharia Elétrica e da Computação) - Universidade Federal de Goiás, Goiânia, 2016.