The most common way to track the position of a vehicle is by using the Global Navigation Satellite System (GNSS). Unfortunately, there are many scenarios where GNSS is inaccessible or provides low precision, and it can therefore be vulnerable to only rely on GNSS. This master's thesis is done in collaboration with the Swedish Defence Research Agency (FOI), who is looking for a solution to this problem. Therefore, this master's thesis develops a system that globally localizes a vehicle in a map, without GNSS. The approach is to combine odometry and the scan registration algorithm iterative closest point (ICP), in an extended Kalman filter (EKF), to provide global position estimates. The ICP algorithm aligns two different sets of data points, referred to as point clouds. In this thesis, one set consists of light detection and ranging (LIDAR) data points collected from a sensor mounted on a vehicle, and the other consists of LIDAR data points collected from an aircraft which forms an elevation map of the area. In the ideal case, the algorithm finds the position on the elevation map where the vehicle collected the data points. For the EKF to function, the uncertainty of ICP must be estimated. Different methods are investigated, which are; unscented transform based covariance, covariance with Hessian, and covariance with correspondences. The result shows that all the methods are too optimistic when estimating the uncertainty. The reason is that none of the methods take all sources of error into account, and it is therefore difficult to correctly capture the uncertainty of ICP. The unscented transform based covariance is the least optimistic, and covariance with correspondences is the most. A second problem investigated in this thesis is how odometry and ICP with an elevation map as reference can be combined to provide a global position estimate. As mentioned, the chosen approach is to implement an EKF which weights the different data sources based on their covariance, to one single estimate. The developed global localization system is evaluated in a real time experiment, where the data is recorded using equipment from FOI. The goal of the experiment is to localize a vehicle while it is driving in different environments, including urban, field and forest environments. The result shows that the performance of the system is viable, and it manages to provide localization within a few meters from ground truth. However, since the ICP covariance estimates are not fully accurate, the performance of the EKF is decreased as it cannot weight the different estimates properly. The ICP algorithm used in the system has a lot of flaws. The worst is that it easily converges to incorrect solutions, in other words that it estimates the wrong position of the vehicle. How this risk can be decreased is also investigated in this thesis. A method that decreases this risk drastically, and makes the viable performance of the system possible, is developed. The approach of the method is to exclude incorrect positions by removing a large amount of points from the point clouds, and keeping the most informative. By only utilizing the most informative data points in the point cloud, global positions with high accuracy are achieved.
Identifer | oai:union.ndltd.org:UPSALLA1/oai:DiVA.org:liu-178858 |
Date | January 2021 |
Creators | Nylén, Rebecka, Rajala, Katherine |
Publisher | Linköpings universitet, Reglerteknik |
Source Sets | DiVA Archive at Upsalla University |
Language | English |
Detected Language | English |
Type | Student thesis, info:eu-repo/semantics/bachelorThesis, text |
Format | application/pdf |
Rights | info:eu-repo/semantics/openAccess |
Page generated in 0.0018 seconds