SLAM landmark identification using lidar measurements

Gericke, Johannes Petrus (2019-04)

Thesis (MEng)--Stellenbosch University, 2019.

Thesis

ENGLISH ABSTRACT: This thesis investigates the use of Lidar sensors for landmark identification in simultaneous localisation and mapping (SLAM). Lidar sensors can very accurately measure the distance toward objects in the environment, but provide no other information about the environment. Modelling landmarks from this spatial representation and associating new measurements with landmarks are important problems to address in order to perform SLAM with these measurements. A literature review shows that multiple different approaches to extracting features from Lidar measurements and modelling landmarks exist. From this study we conclude that existing methods do not reliably associate measurements to landmarks and do not have the ability to update landmark models, which could be helpful to improve the representation of the environment. We consequently develop a probabilistic method to model landmarks from Lidar measurements. In our approach, we approximate object boundaries with continuous, piecewise linear functions. The parameters of these functions are modelled as Gaussian random variables. With this probabilistic model, a method is created to determine if measurements originate from a certain landmark. This method first aligns the measurements to the model using the iterated closest point (ICP) algorithm and then it finds the Mahalanobis distance between measurements and lines to determine if the measurements fit the model. A method is also developed to probabilistically update the model and extend the model when new segments of the landmark are observed. In addition, a SLAM algorithm is designed to use this landmark modelling method. The extended Kalman filter (EKF) SLAM motion update is used directly with an odometry motion model. The measurement update, however, is adapted in order to update the model parameters and robot pose simultaneously. This is in contrast to most other approaches that only use the landmark model to extract a point or reference measurement. Finally, our methods are tested in both simulated environments and with realworld datasets. The results show that the landmark models are good representations of the environment and that measurements of unique objects can be associated with their models. However, the robot tends to be overconfident about its pose and fails to close bigger loops due to faulty associations. We conclude that this approach successfully models the environment with SLAM, but further development needs to take place to make it robust and suitable for practical applications.

AFRIKAANSE OPSOMMING: Hierdie tesis ondersoek die gebruik van Lidar sensors om landmerke te indentifiseer tydens gelyktydige lokalisering en kartering (SLAM). Lidar sensors meet die afstand na voorwerpe in die omgewing met hoë akkuraatheid, maar verskaf geen ander inligting oor die omgewing nie. Die modellering van landmerke met die gebruik van hierdie ruimtelike voorstelling, asook om nuwe metings met landmarke te assosieer, is belangrike probleme om aan te spreek om SLAM uit te voer. ’n Literatuurstudie toon dat verskeie benaderings tot die onttrekking van kenmerke uit Lidar metings, asook die modellering van landmerke bestaan. Uit hierdie studie kom ons tot die gevolgtrekking dat bestaande metodes nie metings betroubaar assosieer met landmerke nie en nie die vermoë het om landmerkmodelle op te dateer nie, wat kan help om die voorstelling van die omgewing te verbeter. Ons ontwikkel dus ’n probabilistiese metode om landmerke met die gebruik van Lidar metings te modelleer. Met hieride metode benader ons objekgrense met kontinue, stuksgewys lineêre funksies. Die parameters van hierdie funksies word gemodelleer as Gaussiese ewekansige veranderlikes. Met die gebruik van hierdie probabilistiese model word ’n metode geskep om vas te stel of nuwe metings van ’n sekere landmerk afkomstig is. Hierdie metode belyn eers die metings met die model met die gebruik van die herhaalde naaste punt (ICP) algoritme en vind dan die Mahalanobis afstand tussen metings en lyne om te bepaal of die metings pas by die model. ’n Metode word ook ontwikkel om die model probabilisties op te dateer en die model uit te brei wanneer nuwe dele van die landmerk waargeneem word. Daarbenewens word ’n SLAM-algoritme ontwerp om hierdie landmerkmodelleringsmetode te gebruik. Die bewegingsopdatering stap van uitgebreide Kalman filter (EKF) SLAM word direk gebruik met ’n odometer-bewegingsmodel. Die meetopdatering stap word egter aangepas sodat die robotposisie en landmerkmodel gelyktydig opgedateer kan word. Dit is in teenstelling met die meeste ander benaderings wat slegs die landmerkmodel gebruik om ’n punt of verwysingsmeting te onttrek. Ten slotte word ons metodes getoets in beide gesimuleerde omgewings en met werklike datastelle. Die resultate toon dat die landmerkmodelle goeie voorstellings van die omgewing skep en dat die metings van unieke voorwerpe met die landmerkmodelle geassosieer kan word. Die robot is egter geneig om te seker die word van sy posisie en versuim om groter lusse in die omgewing te sluit weens foutiewe assosiasies. Ons kom tot die gevolgtrekking dat hierdie metode die omgewing suksesvol modelleer tydens SLAM, maar dat verdere ontwikkeling nodig is om dit robuust en geskik te maak vir praktiese toepassings.

Please refer to this item in SUNScholar by using the following persistent URL: http://hdl.handle.net/10019.1/106197
This item appears in the following collections: