Real Time Lidar and ICP-Based Odometry in Dynamic Environments

Abstract: A robust and highly accurate positioning system is required to transition to fully autonomous vehicles in society.  This thesis investigates the potential for lidar sensors to be a part of a localization system, adding redundancy in case of an outage in a global navigation satellite system GNSS. Point cloud data is recorded on a busy road to experimentally study lidar odometry with dynamic objects present. By matching point clouds with the well-established iterative closest point ICP algorithm, odometry estimates in 6 degrees of freedom are obtained. In this thesis, three ICP variants, point-to-point, point-to-plane and plane-to-plane, are evaluated along with preprocessing and data segmentation techniques to improve accuracy and computational speed.  High-end lidar sensors are known to produce a large amount of data. To achieve real-time performance for the odometry, the point clouds are downsampled using a 3D voxel grid filter to reduce the amount of data by 86% on average. Experiments show that downsampling with a properly tuned voxel grid filter reduces the total process time without sacrificing the accuracy of the estimates. ICP algorithms assume the environment to be static. Therefore dynamic objects can introduce errors in the odometry estimates. Methods to counteract these errors are evaluated. One approach to address this issue, suggested in the literature, is to segment the point cloud into different objects and remove objects smaller than a given threshold. However, experiments on the recorded data set indicate that this method removes too much point cloud data in certain sections, resulting in inaccurate odometry estimates. This problem is especially salient when the environment lacks larger static structures.       However, outlier rejection methods show promising results for suppressing errors caused by dynamic objects. In scan matching, outlier rejection methods can be used to identify and remove individual data point pair associations whose shared distance deviates from the majority in the point clouds. Removing the outliers strengthens the estimates against errors caused by dynamic objects and improves robustness against measurement noise. Experiments in this thesis show that outlier rejection methods can improve translation accuracy with as much as 39% and rotation accuracy with 57% compared to not using any outlier rejection. To improve the accuracy of the estimates, this thesis proposes an approach to divide the lidar point clouds into two subsets, ground points and non-ground points. The scan matching can then be applied to the two subsets separately, enhancing the most relevant information in each subset. Compared to the traditional way of using the entire point clouds in one estimate, experiments show that using the best performing ICP variant, a linearized point-to-plane, in combination with this proposed method improves translation accuracy by 10%, rotation accuracy by 27%, and computational speed by 23%.  The results in this thesis indicate that a lidar odometry solution can be accurate and computationally efficient enough to strengthen a localization system during shorter GNSS outages.