DELO: Deep Evidential LiDAR Odometry using Partial Optimal Transport

Sk Aziz Ali, Djamila Aouada, Gerd Reis, Didier Stricker

Introduction

LiDAR odometry is a critical component for autonomous vehicle navigation, providing 6-degree-of-freedom (6-DoF) motion estimates by comparing sequential laser scans. Despite its importance, current methods struggle with three primary challenges: structural inconsistencies due to non-uniform point densities and partial overlaps, vulnerability to out-of-distribution (OOD) motions like sudden speed changes, and the "black box" nature of deep learning models that lack reliable confidence metrics.

DELO (Deep Evidential LiDAR Odometry) addresses these issues by integrating Partial Optimal Transport (POT) with Evidential Deep Learning. This framework not only estimates motion but also quantifies its own predictive uncertainty, allowing for intelligent, real-time pose-graph optimization.

Methodology

The DELO architecture consists of two main modules: POT-LDM for robust point matching and PU-EPE for uncertainty-aware pose estimation.

Overview of the DELO architecture and its performance on a test sequence.

Partial Optimal Transport (POT-LDM)

Standard matching algorithms often fail by forcing one-to-one correspondences across entire scans, even when points lack a true match in the adjacent frame. DELO utilizes Partial Optimal Transport (POT) to assign higher weights to true inliers while effectively ignoring outliers. This "sharp" matching is implemented via an entropy-regularized Sinkhorn-like algorithm, ensuring that only geometrically consistent points contribute to the motion estimate.

Comparison of matching correspondences between standard DCP and the proposed DELO with Partial Optimal Transport.

Evidential Pose Estimation (PU-EPE)

To move beyond simple point estimates, DELO predicts the parameters of a Normal Inverse-Gamma (NIG) distribution. This approach allows the model to separate aleatoric uncertainty (data noise) from epistemic uncertainty (model ignorance). By quantifying these uncertainties, the system can identify when a pose prediction is likely to be unreliable.

Uncertainty-Aware Pose Refinement

The epistemic uncertainty serves as a functional trigger. When the model's confidence falls outside predefined thresholds, the system automatically initiates a pose-graph optimization using the GTSAM library. This hybrid approach combines the speed of deep learning with the mathematical robustness of classical optimization, performing expensive refinements only when necessary to maintain real-time performance (~35–41ms per frame).

Experimental Results

The framework was evaluated on the KITTI LiDAR Odometry dataset. DELO demonstrated superior generalization, outperforming state-of-the-art methods like PWCLONet on unseen test sequences (07–10).

Key findings include:

  • State-of-the-Art Accuracy: DELO achieved the lowest Relative Translation Error (RTE) and Relative Rotational Error (RRE) across most test sequences.
  • Robustness to OOD Motion: The evidential learning approach proved more resilient to sudden changes in vehicle speed compared to standard supervised methods.
  • Equivariant Uncertainty: The learned uncertainty correctly identifies which specific motion components (e.g., rotation vs. translation) are most prone to error.

Conclusion

DELO presents a significant step forward in LiDAR odometry by making deep learning models "aware" of their own limitations. By leveraging Partial Optimal Transport for robust matching and evidential learning for uncertainty quantification, it achieves a balance of speed, accuracy, and reliability essential for autonomous systems. Future work will explore automated threshold tuning and explicit modeling of dynamic objects to further minimize drift.

Summary updated

Abstract

Accurate, robust, and real-time LiDAR-based odometry (LO) is imperative for many applications like robot navigation, globally consistent 3D scene map reconstruction, or safe motion-planning. Though LiDAR sensor is known for its precise range measurement, the non-uniform and uncertain point sampling density induce structural inconsistencies. Hence, existing supervised and unsupervised point set registration methods fail to establish one-to-one matching correspondences between LiDAR frames. We introduce a novel deep learning-based real-time (approx. 35-40ms per frame) LO method that jointly learns accurate frame-to-frame correspondences and model's predictive uncertainty (PU) as evidence to safe-guard LO predictions. In this work, we propose (i) partial optimal transportation of LiDAR feature descriptor for robust LO estimation, (ii) joint learning of predictive uncertainty while learning odometry over driving sequences, and (iii) demonstrate how PU can serve as evidence for necessary pose-graph optimization when LO network is either under or over confident. We evaluate our method on KITTI dataset and show competitive performance, even superior generalization ability over recent state-of-the-art approaches. Source codes are available.
BibTeX
@inproceedings{ali2023delo,
  author = {Sk Aziz Ali and Djamila Aouada and Gerd Reis and Didier Stricker},
  title = {{DELO: Deep Evidential LiDAR Odometry using Partial Optimal Transport}},
  booktitle = {Workshops @ ICCV},
  year = {2023},
}
Figure