Reconstrução de superfícies 3D a partir de nuvens de pontos usando redes neurais auto-organizáveis

Abstract

This work consists of the implementation of a 3D surface reconstruction algorithm from point clouds, using Self-Organizing Neural Networks. Among the Neural Networks the SOM (Self-Organizing Map) is distinguished, since it is a self-organized model based on competitive learning. The proposed system is based on the GCC (Growing Cell Structures), a self-organizing neural network that is a SOM incremental variation with some alteration in the way that the positions of the direct neighbors of the winner node are updated, and also in the way that new nodes are inserted and inactive nodes are removed. The system has been improved by the inclusion of the edge-swap operation, improving the quality of the generated mesh and the convergence of the algorithm. The algorithm than has been called GCS-M (GCS-Modified). To evaluate the results obtained by the implementation of the proposed system had been used metrics to evaluate the quality of the mesh, based on the values of the minimum distances between the point cloud and the elements that compose the polygon mesh. With this comparison mechanism, it had been performed comparisons of the GCS-M algorithm results and the traditional methods results of surface reconstruction. The obtained results show that the proposed algorithm is very efficient, obtaining realistically reconstructed surfaces with low error measures.