Estratégia de particionamento e pré-alinhamento baseado em distância de Wasserstein para o registro de nuvens de pontos

Abstract

3D point cloud processing is used in multiple areas of interest, such as computer vision, robotics and augmented reality. Likewise, there are many applications that make use of this representation, such as face recognition and automatic navigation systems, among others. However, the benefit of representing objects and scenes in 3D is limited by self-occlusion because, during the acquisition of images in the form of point clouds, the sensor cannot, from a single perspective, capture the entire surface of interest. In general, it is necessary to carry out more than one acquisition in different perspectives, and then combine the views in the same reference in three-dimensional space. This operation is known as point cloud registration, and several techniques have been developed for this purpose. However, registration is a challenging and computationally intensive problem, especially when it comes to finding point correspondences from different regions and the geometric transformations necessary to obtain the desired alignment. The pronounced initial misalignment between partial views can compromise the performance and quality of many recording techniques, and has been the subject of research in many scientific works. In this context, an investigation is carried out with two focuses to help improve registration algorithms: (i) use of the Wasserstein distance to identify correspondence between cloud regions submitted to registration and; (ii) search for a pre-alignment of regions with the highest similarity index to eliminate the problem of the severity of the initial misalignment. As a result, a new pre-alignment technique with low complexity is available that increases the robustness of available registration algorithms in situations of severe misalignment. Experiments based on the Wasserstein distance suggest that it is a measure with potential both for obtaining correspondences between segments of different point clouds and in the pre-alignment process, proving to be a promising approach for future investigations aimed at developing methods that are independent of other algorithms to obtain complete alignment. Future work can also explore the Wasserstein distance to measure the quality of the record in any context, using high similarity correspondences identified in the point clouds presented.