EllipseLIO: Adaptive LiDAR Inertial Odometry with an Ellipsoid Representation

Abstract: LiDAR Inertial Odometry (LIO) is a critical component for many mobile robots that need to navigate without relying on external positioning (e.g., GPS). Platforms that operate autonomously in different environments and with heterogeneous LiDAR sensors require a LIO approach that can adapt to these different scenarios without human intervention. Existing LIO approaches can typically provide reliable and accurate odometry in scenarios with similar environments and sensors when suitably tuned. However, many approaches struggle to retain robust odometry across heterogeneous environments and sensors while using a consistent configuration. This paper presents EllipseLIO, a real-time LIO approach that generalises between scenarios by using methods for LiDAR scan filtering and registration that adapt to the sensor capabilities and environment without requiring scenario-specific tuning. Experiments with EllipseLIO and state-of-the-art LIO approaches on five datasets with diverse and challenging scenarios demonstrate that EllipseLIO is the best-performing approach overall. It achieves a 38% lower odometry error on average than the second-best approach and is the only approach that does not diverge in any experiment. An open-source version of EllipseLIO will be available at github.com/v4rl-ucy/ellipselio.

• The paper presents adaptive LiDAR inertial odometry through range-based scan filtering and ellipsoid registration, eliminating the need for sensor-specific tuning.
• It employs a novel ellipsoid-based registration that integrates point-to-point, point-to-plane, and point-to-line error metrics to robustly capture geometric variations and suppress drift.
• Empirical evaluations show that EllipseLIO reduces absolute pose error by 38% compared to competitors, achieving real-time performance across diverse settings.

Authoritative Summary of "EllipseLIO: Adaptive LiDAR Inertial Odometry with an Ellipsoid Representation" (2605.21150)

Motivation and Problem Statement

LiDAR Inertial Odometry (LIO) is foundational in robotics for GPS-denied navigation, underpinning autonomous operation across platforms and environments. Most extant LIO methods require hand-tuning for each sensor/environment configuration due to fixed scan filtering and rigid error metrics in registration, resulting in poor generalization, divergence, or degraded accuracy when conditions change. EllipseLIO directly confronts this with adaptive mechanisms that eliminate scenario-specific parameterization, targeting robust performance regardless of scene structure or sensor specifications.

Methodological Innovations

Adaptive Range-Based LiDAR Scan Filtering

Traditional LIO approaches downsample scans to fixed voxel resolutions, failing to account for variable sensor resolution, field-of-view, and surface distance. EllipseLIO introduces an algorithm that bins points by radial distance and dynamically sets the voxel resolution per bin based on sensor geometry: closer surfaces retain higher point density, preserving geometric fidelity critical for scan registration, while distant bins are aggressively downsampled to maintain computational efficiency.

This per-bin filtering prevents undesirable sparsity or wasteful density, enabling real-time processing while supporting high-fidelity mapping. The pipeline leverages efficient iOctree structures for insertion/search, making it feasible to run at rates commensurate with modern LiDARs (e.g., 10 Hz).

Ellipsoid-Based Scan Registration

EllipseLIO eschews fixed error metrics in favor of a local surface-adaptive approach. Each map point is associated with an ellipsoid (via Tensor Voting, as per [Medioni2000]) encoding surface variation, decomposed into eigenvalue/eigenvector pairs. Registration for each matched scan-map point uses a convex combination of point-to-point, point-to-plane, and point-to-line metrics, with weights computed directly from saliency measures inferred from the ellipsoid. This enables robust constraints in both structured (planar) and unstructured (ball/line) regions, mitigating the failure modes observed in prior art that assume planarity.

Integrated Drift Correction

Accumulated odometric drift is suppressed via adaptive match weighting—point matches are weighted by trajectory length since the map point was recorded and are modulated for vertical errors (where constraints are typically weaker). Weights are further penalized in degenerate scenarios, using observability heuristics. This correction occurs within the registration pipeline, dispensing with explicit loop closure modules found in SLAM, achieving real-time drift correction during scan/map registration.

Efficient Mapping with Ellipsoid Updates

Mapping aggregates filtered scan points into an iOctree-based global map, with dynamically varying point density and ellipsoid radius per bin. Surface geometry is estimated via multi-stage tensor voting: initial surface normals are refined, and final geometric primitives (plane, line, ball) are assigned via eigenvalue saliency. Updates to ellipsoids are efficient ($O(|Q^\prime|(\log|Q| + |N|))$), parallelizable, and scalable.

Empirical Results

Datasets and Comparative Baselines

EllipseLIO was evaluated on five large-scale, heterogenous datasets spanning urban, botanical, aerial, and waterborne platforms, with varied LiDAR sensors (Ouster OS1-64/OS0-128, Velodyne VLP-16), against state-of-the-art LIO methods: FAST-LIO2, DLIO, LIO-SAM, and iG-LIO. All methods were executed under identical resolutions and hardware environments.

Odometry and Robustness Metrics

The core metric is Absolute Pose Error (APE RMSE), computed after SE(3) alignment with ground truth. EllipseLIO achieves a mean error $38\%$ lower than the next-best method (iG-LIO at $0.5$ m), and, notably, is the only method not to diverge (RMSE $>10$ m) on any sequence, including challenging open-air and dense indoor settings. In structured urban scenarios (Oxford Spires, Newer College), its adaptive registration offers superior or equal performance to planar-centric approaches; in unstructured or aerial datasets, it outperforms all baselines otherwise susceptible to failure modes from inappropriate geometric assumptions or insufficient scan density.

EllipseLIO retains both accuracy and computational efficiency: it processes scans in $35$ ms (parkland-mound), comparable with DLIO at tuned resolutions ($33$–$34$ ms), and uses less memory ($2.4$ GB) to achieve finer map resolution than all other methods except iG-LIO ($0.5$ m, $0.5$ GB), which sacrifices map detail for speed.

Ablation Studies

Key module removal and substitution experiments confirm the essential roles of adaptive scan filtering, ellipsoid registration, and match weighting:

• Removing adaptive match weighting increases mean APE by $38\%$0, due to uncorrected drift.
• Replacing ellipsoid-based registration with point-to-point increases mean APE by $38\%$1, reflecting loss of geometric adaptivity.
• Replacing range-based filtering with uniform $38\%$2 m filtering leads to $38\%$3 APE increase and divergence in large-scale open environments.

All components synergistically contribute to the observed reliability and accuracy.

Theoretical and Practical Implications

The ellipsoid representation fundamentally raises the geometric expressivity of the map, underpinning scan registration by capturing local surface variation far beyond planar assumptions. The adaptive error metric selection, based on locally estimated geometry, represents a shift from fixed-metric registration to context-sensitive registration, offering a substantive generalization to both structured and natural environments.

The drift correction approach, by leveraging revisit weighting in registration rather than explicit loop closures, aligns with recent trends toward lightweight, real-time SLAM systems where computational budgets are constrained.

Practically, EllipseLIO is immediately applicable in heterogeneous robot fleets deployed across varied environments, supporting autonomy with minimal mission-specific setup. The open-source release makes its adoption tractable in both academic and industrial contexts.

Future Directions

Integrating vision with EllipseLIO is the logical extension: visual cues could mitigate degeneracy in geometric environments (e.g., corridors, tunnels) where LiDAR-only odometry is constrained. Further developments may emphasize scalability to multi-agent scenarios, submap partitioning, or efficient semantic mapping. The tensor voting/ellipsoid paradigm introduced here could be generalized to fuse multiple sensor modalities, or to enhance registration with machine-learned geometric priors.

Conclusion

EllipseLIO represents a substantive advance in adaptive LiDAR inertial odometry, offering reliable, sensor- and environment-agnostic performance without scenario-specific tuning. Its range-based filtering, ellipsoid-encoded point representation, adaptive registration error metrics, and integrated drift correction combine to yield robust odometry on all tested datasets, outperforming state-of-the-art methods by significant margins in accuracy and robustness. The theoretical innovations have practical ramifications, enabling deployment across diverse robotic platforms and scenes. Extensions to vision fusion and further geometric abstraction may continue to broaden its impact in navigation and mapping systems.