An essential ability of mobile robots is to build an accurate map of the surrounding environment and localize themselves precisely into it. This allows the robots to be able to operate reliably in complex or unknown environments using only their own perceptions. This is known, in literature, as the Simultaneous Localization and Mapping (SLAM) problem. Over the last decade, many approaches have been proposed to solve the SLAM problem, depending on the type of data gathered by the robot. Particularly accurate and well-known are SLAM systems which use LiDAR data and model the problem as a graph optimization problem (Graph SLAM). Recently, few LiDAR Graph SLAM systems have been proposed, which exploit external maps to insert new information into the graph. These approaches extract features from the external map and associate them with LiDAR perceptions to obtain information about the position of the robot with respect to the features. These approaches are not able to recognize eventual mistakes or misplacements of the features, thus their performance and the correctness of information inserted into the graph could be greatly reduced. In this thesis we introduce three versions of a LiDAR Graph SLAM system, which is able to precisely localize the robot and create an accurate map of the environment, by using information about the surrounding buildings, obtained from maps. The systems are also able, on various degrees, to correct the pose of the buildings that are misplaced into the map. The presented systems are based on an already existing LiDAR Graph SLAM system, called hdl_graph_slam. The first system we propose, called hdl_graph_slam_with_map_priors, introduces a new type of constraint in the graph between a pose of the robot and a building, based on the alignment between the map associated to the buildings and a LiDAR scan. It is able to correct the poses of the robot from drifting but not to correct the poses of the buildings. The second system is called building_slam_with_buildings_priors Rigid SLAM. With respect to the first system (from which it derives), it improves the calculation of the information matrices of the constraints in order to better fit every single building. Differently from the first system, building_slam_with_buildings_priors Rigid SLAM is able to correct also the position of buildings, but not precisely. The last system presented is called building_slam_with_buildings_priors Non-rigid SLAM and constitutes a further improvement of the second system. It improves the calculation of the alignment transformation to adjust perfectly to each building. It is able to reach a good accuracy in the correction of poses of the buildings. To test the proposed systems, we used one trajectory from the KITTI dataset, evaluating both accuracy metrics and visual aspects of the obtained trajectories and maps.
Una capacità essenziale dei robot mobili è quella di costruire una mappa accurata dell'ambiente circostante e di localizzarsi con precisione in essa. Questo permette ai robot di essere in grado di operare in modo affidabile in ambienti complessi o sconosciuti usando solo le loro percezioni. Questo è noto, in letteratura, come problema di Simultaneous Localization and Mapping (SLAM). Nell'ultimo decennio sono stati proposti molti approcci per risolvere il problema dello SLAM, a seconda del tipo di dati raccolti dal robot. Particolarmente precisi e conosciuti sono i sistemi di SLAM che usano dati provenienti da un LiDAR e modellano il problema come un problema di ottimizzazione del grafo (Graph SLAM). Recentemente, sono stati proposti alcuni sistemi di LiDAR Graph SLAM che sfruttano mappe esterne per inserire nuove informazioni nel grafo. Questi approcci estraggono caratteristiche dalla mappa esterna e le associano alle percezioni del LiDAR per ottenere informazioni sulla posizione del robot rispetto a tali caratteristiche. Questi approcci non sono in grado di riconoscere eventuali errori o posizionamenti errati delle caratteristiche, quindi le loro prestazioni e la correttezza delle informazioni inserite nel grafo potrebbero essere notevolmente ridotte. In questa tesi presentiamo tre versioni di un sistema di LiDAR Graph SLAM, che è in grado di localizzare con precisione il robot e creare una mappa accurata dell'ambiente, utilizzando le informazioni sugli edifici circostanti, ottenute dalle mappe. I sistemi sono anche in grado, a vari gradi, di correggere le pose degli edifici che sono mal posizionati nella mappa. I sistemi presentati sono basati su un sistema di LiDAR Graph SLAM già esistente, chiamato hdl_graph_slam. Il primo sistema che proponiamo, chiamato hdl_graph_slam_with_map_priors, introduce un nuovo tipo di vincolo nel grafo tra una posa del robot e un edificio, basato sull'allineamento tra la mappa associata agli edifici e una scansione LiDAR. È in grado di correggere le pose del robot da imprecisioni, ma non di correggere le pose degli edifici. Il secondo sistema è chiamato building_slam_with_buildings_priors Rigid SLAM. Rispetto al primo sistema (da cui deriva), migliora il calcolo delle matrici di informazione dei vincoli per adattarsi meglio ad ogni singolo edificio. A differenza del primo sistema, building_slam_with_buildings_priors Rigid SLAM è in grado di correggere anche le posizioni degli edifici, ma non con precisione. L'ultimo sistema presentato è chiamato building_slam_with_buildings_priors Non-rigid SLAM e costituisce un ulteriore miglioramento del secondo sistema. Migliora il calcolo della trasformazione di allineamento per adattarsi perfettamente ad ogni edificio. È in grado di raggiungere una buona precisione nella correzione delle pose degli edifici. Per testare i sistemi proposti, abbiamo utilizzato una traiettoria dal dataset KITTI, valutando sia le metriche di precisione che gli aspetti visivi delle traiettorie e delle mappe ottenute.
Simultaneous localization and mapping in urban scenarios with non-rigid OpenStreetMap priors
Gobbi, Veronica
2020/2021
Abstract
An essential ability of mobile robots is to build an accurate map of the surrounding environment and localize themselves precisely into it. This allows the robots to be able to operate reliably in complex or unknown environments using only their own perceptions. This is known, in literature, as the Simultaneous Localization and Mapping (SLAM) problem. Over the last decade, many approaches have been proposed to solve the SLAM problem, depending on the type of data gathered by the robot. Particularly accurate and well-known are SLAM systems which use LiDAR data and model the problem as a graph optimization problem (Graph SLAM). Recently, few LiDAR Graph SLAM systems have been proposed, which exploit external maps to insert new information into the graph. These approaches extract features from the external map and associate them with LiDAR perceptions to obtain information about the position of the robot with respect to the features. These approaches are not able to recognize eventual mistakes or misplacements of the features, thus their performance and the correctness of information inserted into the graph could be greatly reduced. In this thesis we introduce three versions of a LiDAR Graph SLAM system, which is able to precisely localize the robot and create an accurate map of the environment, by using information about the surrounding buildings, obtained from maps. The systems are also able, on various degrees, to correct the pose of the buildings that are misplaced into the map. The presented systems are based on an already existing LiDAR Graph SLAM system, called hdl_graph_slam. The first system we propose, called hdl_graph_slam_with_map_priors, introduces a new type of constraint in the graph between a pose of the robot and a building, based on the alignment between the map associated to the buildings and a LiDAR scan. It is able to correct the poses of the robot from drifting but not to correct the poses of the buildings. The second system is called building_slam_with_buildings_priors Rigid SLAM. With respect to the first system (from which it derives), it improves the calculation of the information matrices of the constraints in order to better fit every single building. Differently from the first system, building_slam_with_buildings_priors Rigid SLAM is able to correct also the position of buildings, but not precisely. The last system presented is called building_slam_with_buildings_priors Non-rigid SLAM and constitutes a further improvement of the second system. It improves the calculation of the alignment transformation to adjust perfectly to each building. It is able to reach a good accuracy in the correction of poses of the buildings. To test the proposed systems, we used one trajectory from the KITTI dataset, evaluating both accuracy metrics and visual aspects of the obtained trajectories and maps.File | Dimensione | Formato | |
---|---|---|---|
2021_07_GOBBI_def.pdf
accessibile in internet solo dagli utenti autorizzati
Dimensione
8.98 MB
Formato
Adobe PDF
|
8.98 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/177170