D-LIO

The method is based on the proposed Fast Truncated Distance Field (Fast-TDF) as a convenient tool to represent the environment. This representation enables solving the LiDAR point-cloud registration as a nonlinear optimization process, simultaneously producing an accurate TDF map of the environment, and updating such map at constant time independently of its size. The approach is benchmarked against other state-of-the-art odometry approaches, demonstrating the same or better level of accuracy.

The D-LIO algorithm incrementally estimates the system pose using an Extended Kalman Filter (EKF). It integrates inertial measurements for pose prediction and uses point cloud-to-map registration for pose updates. The environment is mapped based on the Truncated Distance Field (TDF). This alignment is used to constrain the EKF's pose and velocity. Whenever the estimated pose variation exceeds a threshold, the TDF map is updated by incorporating new information.

Figure 1: The final distance map (center) is computed by applying the bitwise-AND operation to overlapping kernels (top/bottom).

We benchmarked D-LIO on three challenging public datasets: VIRAL, Newer College, and VBR. These tests evaluate performance in diverse scenarios, including aerial (UAV) and ground platforms, against several state-of-the-art (SOTA) methods.

Trajectory Accuracy (ATE)

The Absolute Translation Error (ATE) [m] was measured on the VIRAL and Newer College datasets. As shown in the table, D-LIO achieves consistently high accuracy, ranking first or second in most sequences. It remains highly competitive even when compared to methods that utilize back-end optimization (like LIO-SAM or PIN-SLAM), which D-LIO does not.

Table 2: ATE (m) on VIRAL and Newer College. Best (bold) and second-best (underline) odometry methods.

Scalability on Large Datasets

To evaluate performance in large-scale environments, we used the VBR dataset, featuring long trajectories (up to 2.5 km) in complex indoor/outdoor settings. D-LIO demonstrates robust scalability, maintaining low ATE (m) and Relative Pose Error (RPE) (%).

Table 3: ATE (m) and RPE (%) on the VBR dataset.

---

Mapping Quality

We also evaluated the 3D reconstruction quality on the Newer College (Quad-Easy) sequence using standard metrics: Accuracy, Completeness, Chamfer-L1 (C-L1), and F-Score. D-LIO produces highly competitive maps, achieving the best completeness and ranking second for C-L1 and F-Score among all tested odometry-driven methods.

Table 4: Mapping metrics on Newer College. Best (bold) and second-best (underline) scores.

Runtime Analysis

The system's performance was analyzed on an Intel Core i7-13700H CPU without GPU acceleration. The map update step is the most computationally expensive, as its time strongly depends on point cloud density. However, optimization runs at the sensor rate (10 Hz) while updates are triggered less frequently, balancing the load.

Table 5: Runtime analysis (mean ± std. dev) with no downsampling.

Project Video

Qualitative results, 3D reconstructions, and estimated trajectories from the datasets.

If you use D-LIO, please cite:

@article{cotoelena2025dlio,
  title   = {D-LIO: 6DoF Direct LiDAR-Inertial Odometry based on Simultaneous Truncated Distance Field Mapping}, 
  author  = {Lucia Coto-Elena and J. E. Maese and L. Merino and F. Caballero},
  journal = {IEEE Robotics and Automation Letters (RA-L)},
  year    = {2025},
  note    = {Accepted for publication. To appear.},
  eprint  = {2505.16726},
  archivePrefix = {arXiv},
  url     = {https://arxiv.org/abs/2505.16726}
}

This work is partially supported by the grants INSERTION (PID2021- 127648OB-C31) and NORDIC (TED2021-132476B-I00), both funded by the “Agencia Estatal de Investigación – Ministerio de Ciencia, Innovación y Universidades” and the “European Union NextGenerationEU/PRTR.