DGV-DLIVO: Degeneracy-Guided Visual Fusion for Direct LiDAR-Inertial-Visual Odometry in Degenerate Tunnel Environments

Junrui ZhangarXiv 2026 · arXiv

Reliable localization is the "holy grail" for autonomous robots in underground environments. While LiDAR-Inertial Odometry (LIO) has become a standard for many applications, it faces a fundamental physics problem in long, featureless tunnels: structural degeneracy. In a perfectly smooth, straight pipe, moving forward looks identical to standing still to a LiDAR sensor. This "infinite corridor problem" leads to stagnation, where the robot's estimated position stops progressing even as it moves physically.

A new paper from researchers at Tongji University introduces DGV-DLIVO, a framework designed to break this deadlock by using vision not just as an extra sensor, but as a targeted "fix" for LiDAR’s blind spots.

TL;DR

DGV-DLIVO solves the problem of LiDAR "stagnation" in long tunnels by using the detected direction of geometric degeneracy to guide visual feature selection and fusion. By selectively injecting visual constraints only into the motion subspaces where LiDAR is weak, it achieves a 91.4% progress ratio in extreme hydropower tunnel environments where state-of-the-art methods fail.

What Problem Does This Paper Address?

In hydropower spillway tunnels and drainage galleries, two things happen simultaneously to break traditional odometry:

  1. Geometric Degeneracy: The tunnel walls are so uniform that LiDAR scan matching cannot distinguish between different longitudinal positions. The "constraints" along the tunnel axis are effectively zero.
  2. Visual Degradation: These environments are often dimly lit with repetitive, weak textures (bare concrete), making standard Visual-Inertial Odometry (VIO) prone to drift or divergence.

Existing "tightly coupled" systems (like R3LIVE or FAST-LIVO2) often treat vision as a general-purpose supplement. However, in these extreme cases, uncertain visual data can actually "pollute" the reliable geometric constraints LiDAR does provide (like lateral position and heading), leading to erratic trajectories.

The challenge: A long, feature-scarce hydropower tunnel and its reconstructed map.

Methods

The core innovation of DGV-DLIVO is Degeneracy-Guided Visual Fusion. Instead of a "more is better" approach to data, the system follows a targeted pipeline:

  1. Degeneracy Detection: The system analyzes the LiDAR information matrix to identify the "weak direction" ($d^W$)—the axis along which the environment provides the least geometric information.
  2. Guided Patch Discovery: Rather than looking for generic corners, the system selects visual patches that have strong gradients aligned with the LiDAR-degenerate direction. If the LiDAR can't see "forward," the camera specifically looks for visual cues (stains, joints, markings) that help define "forward" motion.
  3. LiDAR-Anchored Association: Visual patches are "anchored" to the metric depth provided by the LiDAR map. This turns 2D image pixels into 3D constraints without needing complex monocular triangulation.
  4. Subspace-Selective Fusion: This is the mathematical "secret sauce." The system projects visual residuals only into the deficient motion subspace. Simultaneously, it adaptively suppresses unreliable LiDAR components in that same direction while preserving the "clean" LiDAR data that constrains lateral and vertical motion.
System overview showing the interplay between the LiDAR-Inertial stream and the Degeneracy-Guided Visual block.

Datasets

The authors evaluated their system in real-world hydropower spillway tunnels. Because standard public datasets (like KITTI or UrbanNav) do not feature this specific type of extreme longitudinal degeneracy, the researchers collected their own data using a custom platform equipped with a Livox Mid360 LiDAR and a Basler camera.

  • Cart-mounted sequence: A 40m run with sparse wall markers used as ground-truth references.
  • Handheld sequence: An oblique-view run designed to test robustness against aggressive motion and varied viewing angles.

Key Insights and Ideas

  • Active Guidance vs. Passive Fusion: Most fusion algorithms are passive—they take what the sensors give them. DGV-DLIVO uses the failure of one sensor (LiDAR) as an active signal to tell the other sensor (Camera) where to focus its "attention."
  • Directional Weighting: As shown in Figure 3, the system generates a "direction-of-interest" vector. This ensures the visual system isn't wasting computational power on features that don't help solve the degeneracy.
  • Subspace Masking: The realization that "bad" data is worse than "no" data. By explicitly suppressing the noisy LiDAR components along the tunnel axis, the system prevents the "stiffness" that usually causes LIO systems to stagnate.
Visual patch selection process: (a) input image, (b) direction-weighted response, and (e) the final selected patches.

Key Results

In the 40m tunnel test, DGV-DLIVO was the only method to maintain continuous, directionally consistent motion:

  • Progress Ratio: Achieved 91.4%, meaning it successfully recovered almost the entire length of the tunnel.
  • Comparison: Standard LIO methods (FAST-LIO2, DLIO) essentially "stuck" at the start (0.1% to 10% progress). Tightly coupled LIVO methods (R3LIVE) managed about 58% progress but suffered from significant "under-travel."
  • Stability: While VINS-Mono (VIO) could estimate the distance, its trajectory suffered from severe "curved drift" due to the lack of geometric anchors. DGV-DLIVO remained straight and stable.
Trajectory comparison: DGV-DLIVO (red) follows the ground truth markers, while others stagnate or drift.

Limitations and Future Work

While highly effective, the authors note some areas for improvement:

  • Texture Requirements: The system still requires some visual information. In a perfectly pitch-black, polished pipe with zero texture, even this method would eventually fail.
  • Computational Load: While designed to be lightweight (running on an i7 CPU without a discrete GPU), the patch-based ZNCC (Zero-mean Normalized Cross-Correlation) search adds overhead compared to pure LIO.
  • Future Directions: The authors suggest exploring more sophisticated "full-attention" modes where visual information can act on the local LiDAR tangent plane more broadly, and potentially integrating loop closure for even longer-range consistency.

For those interested in the implementation, the source code is available on GitHub.

Summary updated

Abstract

Reliable localization in long underground infrastructures remains a major challenge for autonomous inspection robots. In hydropower spillway tunnels and drainage galleries, repetitive geometric structures provide insufficient constraints along the tunnel axis, while weak texture and illumination variations further degrade visual estimation. Under this coupled degradation, existing LiDAR-inertial and visual-inertial odometry systems may lose the ability to recover continuous longitudinal motion. This paper presents DGV-DLIVO, a degeneracy-guided visual fusion for direct LiDAR-inertial-visual odometry for extreme tunnel environments. Instead of treating vision as a general complementary sensor, the proposed method exploits the LiDAR-degenerate direction as an active guidance signal to determine informative visual regions, establish LiDAR-anchored visual associations, and perform subspace-selective fusion. The visual constraints are introduced only into the deficient motion subspace, while unreliable LiDAR information in the same direction is adaptively suppressed and reliable geometric constraints are preserved. Real-world experiments in hydropower tunnels demonstrate that representative state-of-the-art LIO, LIVO, and VIO methods suffer from stagnation, oscillation, severe under-travel, curved drift, or divergence, whereas DGV-DLIVO maintains continuous and directionally consistent tunnel-axis motion estimation under both cart-mounted and handheld acquisition. It achieves a 91.4% progress ratio over the 40m tunnel sequence and significantly improves longitudinal recovery and trajectory stability. These results demonstrate the effectiveness of degeneracy-guided adaptive fusion for robotic inspection in severely degraded underground environments. The source code and a supplementary video are publicly available at https://github.com/0RayZhang0/DGV-DLIVO.
BibTeX
@misc{zhang2026dgv,
  author = {Junrui Zhang},
  title = {{DGV-DLIVO: Degeneracy-Guided Visual Fusion for Direct LiDAR-Inertial-Visual Odometry in Degenerate Tunnel Environments}},
  year = {2026},
  eprint = {2611.22385},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.22385},
}
Figure