Modeling Environments Hierarchically with Omnidirectional Imaging and Global-Appearance Descriptors

Abstract

Abstract:In this work, a framework is proposed to build topological models in mobile robotics, using an omnidirectional vision sensor as the only source of information. The model is structured hierarchically into three layers, from one high-level layer which permits a coarse estimation of the robot position to one low-level layer to refine this estimation efficiently. The algorithm is based on the use of clustering approaches to obtain compact topological models in the high-level layers, combined with global appearance techniques to represent robustly the omnidirectional scenes. Compared to the classical approaches based on the extraction and description of local features, global-appearance descriptors lead to models that can be interpreted and handled more intuitively. However, while local-feature techniques have been extensively studied in the literature, global-appearance ones require to be evaluated in detail to test their efficacy in map-building tasks. The proposed algorithms are tested with a set of publicly available panoramic images captured in realistic environments. The results show that global-appearance descriptors along with some specific clustering algorithms constitute a robust alternative to create a hierarchical representation of the environment.

A Novel Method to Estimate the Position of a Mobile Robot in Underfloor Environments Using RGB-D Point Clouds

“…, 36. To determine this correspondence, global appearance descriptors have been used to extract the most relevant information from the images [44]- [46], to compare them pairwise and to make the pairing process. In particular, the Fourier Signature (FS), proposed initially by Menegatti et al [47] has been used.…”

RGB-D Object SLAM Using Quadrics for Indoor Environments

“…The conventional visual SLAM extracts static geometric features in the environment, such as points, lines, and planes, and achieves high-precision localization and mapping [ 1 ]. Some also use the global appearance of whole images such as global-appearance descriptors [ 2 , 3 ], or gist descriptor [ 4 ] for mapping or localization, which model more information in an image compared with features.…”

Uso de técnicas de machine learning para realizar mapping en robótica móvil

“…Estos trabajos se basan en el uso de descriptores que se calculan de manera analítica, sin embargo, en loś ultimos años ha emergido el uso de descriptores de apariencia global basados en técnicas de deep learning. Por ejemplo, Xu et al [14] propuso de auto-encoders para obtener descriptores de apariencia global que le permitiesen detectar tumores de pecho; Payá et al [10] utiliza los vectores obtenidos de capas intermedias de una red neuronal convolucional (CNN) como descriptores de apariencia global y a partir de esta información lleva a cabo la creación de mapas jerárquicos.…”

“…MAPPINGAdemás de las herramientas de machine learning mencionadas, en este trabajo también se estudia el uso de descriptores de apariencia global, los cuales han sido utilizados durante losúltimos años para resolver la tarea de mapping. Por ejemplo, Payá et al[10] utilizaron descripción de apariencia global para construir modelos topológicos jerárquicos. Basándonos en este trabajo, el presente estudio analiza el uso de los descriptores HOG, gist y un descriptor basado en deep learning.…”