Outlier-Robust Long-Term Robotic Mapping Leveraging Ground Segmentation

TL;DR

Addressing the brittleness of learning-based perception in long-term robotic mapping, the paper advocates a robust, conventional approach for diverse environments. It introduces a learning-free pipeline with fast ground segmentation, GNC-based outlier-robust registration for loop closing and initial multi-session alignment, hierarchical multi-session SLAM, and instance-aware static map building to handle moving objects. The approach leverages ground segmentation to prune ground points, applies GNC for robust pose estimation amid large viewpoint differences, and removes dynamic points to produce static maps for localization and planning, reducing false loops and improving map quality. This combination supports reliable loop closure, cross-session alignment, and long-term autonomy across varied robots and sensors without reliance on data-driven models that may fail in unmodeled scenes.

Abstract

Despite the remarkable advancements in deep learning-based perception technologies and simultaneous localization and mapping (SLAM), one can face the failure of these approaches when robots encounter scenarios outside their modeled experiences (here, the term modeling encompasses both conventional pattern finding and data-driven approaches). In particular, because learning-based methods are prone to catastrophic failure when operated in untrained scenes, there is still a demand for conventional yet robust approaches that work out of the box in diverse scenarios, such as real-world robotic services and SLAM competitions. In addition, the dynamic nature of real-world environments, characterized by changing surroundings over time and the presence of moving objects, leads to undesirable data points that hinder a robot from localization and path planning. Consequently, methodologies that enable long-term map management, such as multi-session SLAM and static map building, become essential. Therefore, to achieve a robust long-term robotic mapping system that can work out of the box, first, I propose (i) fast and robust ground segmentation to reject the ground points, which are featureless and thus not helpful for localization and mapping. Then, by employing the concept of graduated non-convexity (GNC), I propose (ii) outlier-robust registration with ground segmentation that overcomes the presence of gross outliers within the feature matching results, and (iii) hierarchical multi-session SLAM that not only uses our proposed GNC-based registration but also employs a GNC solver to be robust against outlier loop candidates. Finally, I propose (iv) instance-aware static map building that can handle the presence of moving objects in the environment based on the observation that most moving objects in urban environments are inevitably in contact with the ground.

Figures (4)

• Figure 1: (a)-(b) Performance comparison between Patchwork++ lee2022patchworkpp and our proposed approach, which rejects the reflected noises and thus successfully filters ground points. Light green and red bins represent that normal vectors (arrows) of the estimated planes that are sufficiently upright and too tilted, respectively. Blue points denote wrongly rejected ground points, i.e., false negatives (best viewed in color).
• Figure 2: (a) Example of source (red) and target clouds (green) in the KITTI dataset, where the left-bottom text represents the pose discrepancy in each scene. (b)-(c) Registration results before (W/o GS) and after the application of ground segmentation (W/ GS), where the blue points denote the warped source cloud by the estimated pose. Our proposed approach lim2023quatro++ successfully estimates the relative pose even though the pose discrepancy between viewpoints of source and target clouds is far distant. Black dashed boxes highlight the non-ground objects that have to be tightly aligned (note that, in (c), these objects are well-aligned). The solid red and green boxes indicate that algorithms failed and succeeded, respectively (best viewed in color).
• Figure 3: Our proposed long-term mapping pipeline. (a) Registration result of our outlier-robust registration framework when estimating the relative pose between two distant viewpoints of point clouds. Our approach robustly estimates the relative pose despite gross outliers (red lines). (b)-(c) Before and after the application of our proposed registration in the loop closing module of SC-LeGO-LOAM kim2018scancontext. Our proposed registration significantly reduces the number of false positive loops, enhancing the mapping result (see the upper zoomed boxes). (d) Our multi-session SLAM results by taking two trajectories from our single-session SLAM as inputs. Note that we leverage our outlier-robust registration to initially align the data via map-to-map registration. (e) Example of results by the proposed static map building approach. Because the points from moving objects are rejected at an instance-level, traces of pedestrians and vehicles are successfully filtered out. Note that all the approaches are designed to be learning-free and robust; thus, the proposed approaches can be easily employed in various robot platforms and sensor configurations.
• Figure 4: Ongoing research topics: (a) semantics-aware registration output in the MulRan dataset kim2020mulran using the semantic segmentation network trained in the SemanticKITTI dataset behley2019iccv and (b) scene graph-based mapping result in our campus with zero-shot predictions sun2019high.