Towards lifelong slam: robust loop closure detection and verification over time

• Autores: Yasir Latif
• Directores de la Tesis: José Neira Parra (dir. tes.)
• Lectura: En la Universidad de Zaragoza ( España ) en 2014
• Idioma: español
• Tribunal Calificador de la Tesis: Juan Domingo Tardos Solano (presid.), Rafael García Campos (secret.), Davide Scaramuzza (voc.)
• Texto completo no disponible (Saber más ...)
• Resumen

  • Answering the question: ¿Where am I?¿ is a difficult task for a mobile robot. In order to know its precise location, it needs to build a coherent representation of the operating environment. However, building such a representations requires correctly knowing the position of the robot. Therefore, algorithms that address this problem solve for both these unknowns at the same time and are termed Simultaneous Localization And Mapping (SLAM) algorithms.

    An important aspect of the SLAM formulation is place recognition, or loop closure detection, and this is what differentiates SLAM from sensor based odometry. It is of utmost important to correctly close loops as it bounds the otherwise unbounded uncertainty in the robot¿s location and improves the precision of the map estimate. However, loop closure detection algorithm are prone to perceptual aliasing due to self similarities in the environment, which can lead to corruption of the environment representation being estimated. This work presents two contributions towards the problem of place recognition: a) By realizing that loop closures occur sparsely, i.e, the current place corresponds to only a few probable candidate locations from the past, the problem of loop closure detection can be cast as a sparse optimization problem. This formulation allows certain desirable qualities such as global uniqueness and flexibility of representation. b) The second contribution of this work deals with detecting and removing inconsistent loop closures arising from perceptual aliasing which would otherwise irreparably corrupt the map estimate. We propose a ¿Robust SLAM backend¿ algorithm that tries to find the maximally consistent subset of loop closures based on a series of consistency tests. The algorithm is designed to reconsider decisions made at each step, allowing it to reverse decisions as new evidence becomes available, effecting reasoning over time to select the set of loop closing decisions that is consistent with a given robot trajectory. Furthermore, we also propose a method for reducing the number of poses present in a pose graph formulation of the SLAM problem so that algorithms that do not need the whole graph for reasoning can benefit from the reduction in complexity but still enjoy the same accuracy as the complete system.