Ground Level Propagation: A Novel Moving Object Segmentation Component to Process Ground in Multi-Level Structures

Abstract

Moving Object Segmentation (MOS) in LiDAR point clouds is an essential component of autonomous driving and map cleaning applications. Existing ground segmentation algorithms - which are used as a foundational component of many MOS frameworks - fail in scenes with multi-level structures like bridges and overpasses, because they rely on the assumption that there is only a single ground level per vertical column in the point cloud. This assumption significantly reduces segmentation quality in scenes with more than one ground level.

This thesis introduces a novel offline MOS framework with a component specifically designed to address this multi-level issue, which is our primary contribution. This component, called Ground Level Propagation (GLP), is a multi-level ground segmentation module built upon the plane fitting approach used by Patchwork++. It iteratively identifies and fits ground planes across unsegmented areas. This results in stable performance in complex scenes, and provides a natural way of handling below-ground noise points.

The proposed framework identifies dynamic points by combining the speed of a visibility-based approach (BeautyMap) with the flexibility of a probabilistic approach (ERASOR2), allowing the user to easily adjust the precision-recall trade-off. We also discuss several smaller contributions, including a cylindrical ray cast, an alternative method for updating voxel statuses in the occlusion map, and probability updates based on the distance to the LiDAR scanner.

The performance of the framework was tested by a large number of experiments: hyperparameter optimization, an ablation study on the SemanticKITTI dataset, qualitative analysis on a private dataset with many multi-level structures, and a comparison to state-of-the-art MOS methods. The results of these experiments show that the GLP component provides a major improvement over current state-of-the-art ground segmentation methods in multi-level structures. However, the experiments also indicate weaker generalizability across different ego-vehicle speeds compared to learning-based approaches. Overall, with the newly proposed GLP component, this research provides a solid solution for ground segmentation in scenes featuring multi-level structures, and can serve as a foundation for future MOS methods that aim to extend their segmentation coverage to such scenes.