Avaliação de algoritmo de clusterização para segmentação de nuvens de pontos 3D

Abstract

3D technologies have been explored in many industrial fields. With those technologies functions such as visualization, instrumentation, control, simulation, training, planning, documentation, and more are performed. From that, new types of media were introduced to the industrial context. For instance, the point clouds, which is a set of points distributed on a tridimensional model of reality. They are often built through the operation of a scanner and may contain the characteristics of an object, such as localization, color, reflectance and more in every one of its points. Given that the point clouds represent pieces, equipments, piping, machines, areas and structures in the industrial context, a useful tool is to segment the point clouds and enable a better visualization of separate parts of the same structure. Furthermore, among the presented 3D technologies, point cloud segmentation permeates directly and indirectly the fields of visualization, instrumentation, and control. Thus, verified the importance of the point cloud segmentation, the goal of this monography is to assess two clustering algorithms for 3D point cloud segmentation. They are the DBSCAN and the K-means. Both are categorized as clustering algorithms through unsupervised machine learning. After the assessment, the differences between each algorithm were found. The K-means was verified to have better performance when it comes to disperse data and the equivalent to the DBSCAN when it comes to data distribution with a certain distance between the clusters.