Detecção de mudanças e recuperação de forma em mapas 3D basedos em nuvens de pontos

Abstract

This work proposes a framework to efficiently detect and segment changes in 3D point clouds. Initially, we simplify the point cloud which generates a tree representation in order to preserve the relevant characteristics using surface variation as the key constraint. Then, a combined approach using Gaussian Mixture Models (GMM) and the Earth Mover's Distance (EMD) is proposed, together with a greedy technique. It detects what has changed in a given locale from a previous sensor reading and segments the detected novelty.The 3D shape retrieval is achieved using two different approaches. The first approach is a new method that we developed which directly retrieves the detected shape in the Gaussian space. It uses a limited set or geometric primitives including planes, spheres and cylinders. This method is compared with another one that operates in the Euclidean space using Random Sampling Consensus (RANSAC). The second approach uses superquadric models due to their good expressiveness in representing more generalized three dimensional shapes. In order to achieve a good level of efficiency and to overcome the limitations of GMM segmentation, a multi-scale approach using the split-and-merge paradigm was devised and successfully implemented.Experimental results in both real and simulated scenarios were obtained to evaluate the efficacy of the simplification and change detection methods. The shape retrieval method in the Gaussian space showed better results than the Euclidean space method, both in accuracy and computational cost. The limitations on shape representation were adequately overcome by the use of superquadrics.