Data-driven covariance estimation for the iterative closest point algorithm
|Advisor:||Giguère, Philippe; Pomerleau, François|
|Abstract:||Three-dimensional point clouds are an ubiquitous data format in robotics. They are produced by specialized sensors such as lidars or depth cameras. The point clouds generated by those sensors are used for state estimation tasks like mapping and localization. Point cloud registration algorithms, such as Iterative Closest Point (ICP), allow us to make ego-motion measurements necessary to those tasks. The fusion of ICP registrations in existing state estimation frameworks relies on an accurate estimation of their uncertainty. Unfortunately, existing covariance estimation methods often scale poorly to the 3D case. This thesis aims to estimate the uncertainty of ICP registrations for 3D point clouds. First, it poses theoretical foundations from which we can articulate a covariance estimation method. It reviews the ICP algorithm, with a special focus on the parts of it that are pertinent to covariance estimation. Then, an inserted article introduces CELLO-3D, our data-driven covariance estimation method for ICP. The article contains a thorough experimental validation of the new algorithm. The latter is shown to perform better than existing covariance estimation techniques in a wide variety of environments. Finally, this thesis comprises supplementary experiments, which complement the article.|
|Document Type:||Mémoire de maîtrise|
|Open Access Date:||6 May 2019|
|Collection:||Thèses et mémoires|
All documents in CorpusUL are protected by Copyright Act of Canada.