Location models for visual place recognition

Elena Stumm 1
1 LAAS-RIS - Équipe Robotique et InteractionS
LAAS - Laboratoire d'analyse et d'architecture des systèmes [Toulouse]
Abstract : This thesis deals with the task of appearance-based mapping and place recognition for mobile robots. More specifically, this work aims to identify how location models can be improved by exploring several existing and novel location representations in order to better exploit the available visual information. Appearance-based mapping and place recognition presents a number of challenges, including making reliable data-association decisions given repetitive and self-similar scenes (perceptual aliasing), variations in view-point and trajectory, appearance changes due to dynamic elements, lighting changes, and noisy measurements. As a result, choices about how to model and compare observations of locations is crucial to achieving practical results. This includes choices about the types of features extracted from imagery, how to define the extent of a location, and how to compare locations. Along with investigating existing location models, several novel methods are developed in this work. These are developed by incorporating information about the underlying structure of the scene through the use of covisibility graphs which capture approximate geometric relationships between local landmarks in the scene by noting which ones are observed together. Previously, the range of a location generally varied between either using discrete poses or loosely defined sequences of poses, facing problems related to perceptual aliasing and trajectory invariance respectively. Whereas by working with covisibility graphs, scenes are dynamically retrieved as clusters from the graph in a way which adapts to the environmental structure and given query. The probability of a query observation coming from a previously seen location is then obtained by applying a generative model such that the uniqueness of an observation is accounted for. Behaviour with respect to observation errors, mapping errors, perceptual aliasing, and parameter sensitivity are examined, motivating the use of a novel normalization scheme and observation likelihoods representations. The normalization method presented in this work is robust to redundant locations in the map (from missed loop-closures, for example), and results in place recognition which now has sub-linear complexity in the number of locations in the map. Beginning with bag-of-words representations of locations, location models are extended in order to include more discriminative structural information from the covisibility map. This results in various representations ranging between unstructured sets of features and full graphs of features, providing a tradeoff between complexity and recognition performance.
Document type :
Robotics [cs.RO]. Universite Toulouse III Paul Sabatier, 2015. English
Liste complète des métadonnées

Contributor : Arlette Evrard <>
Submitted on : Monday, November 7, 2016 - 1:47:59 PM
Last modification on : Thursday, January 11, 2018 - 6:26:55 AM
Document(s) archivé(s) le : Wednesday, February 8, 2017 - 12:02:08 PM


  • HAL Id : tel-01376134, version 1


Elena Stumm. Location models for visual place recognition. Robotics [cs.RO]. Universite Toulouse III Paul Sabatier, 2015. English. 〈tel-01376134〉



Record views


Files downloads