[Paper Review] 3D Surface Reconstruction from Voxel-based Lidar Data

Luis Roldão, Raoul de Charette+1 more|arXiv (Cornell University)|Jun 25, 2019

Robotics and Sensor-Based Localization30 references21 citations

TL;DR

This paper proposes a 3D surface reconstruction method from heterogeneous-density Lidar data using a TSDF voxel representation with an adaptive neighborhood kernel and Gaussian confidence estimation. By leveraging local planar surface approximations and confidence-aware neighborhood selection, the method achieves higher reconstruction accuracy than IMLS while maintaining sufficient mesh density near the sensor, outperforming state-of-the-art methods on both CARLA and KITTI datasets.

ABSTRACT

To achieve fully autonomous navigation, vehicles need to compute an accurate model of their direct surrounding. In this paper, a 3D surface reconstruction algorithm from heterogeneous density 3D data is presented. The proposed method is based on a TSDF voxel-based representation, where an adaptive neighborhood kernel sourced on a Gaussian confidence evaluation is introduced. This enables to keep a good trade-off between the density of the reconstructed mesh and its accuracy. Experimental evaluations carried on both synthetic (CARLA) and real (KITTI) 3D data show a good performance compared to a state of the art method used for surface reconstruction.

Motivation & Objective

• Address the challenge of reconstructing accurate 3D surfaces from noisy, sparsely sampled, and heterogeneously dense Lidar point clouds.
• Improve the trade-off between mesh accuracy and density in 3D surface reconstruction, especially near the sensor.
• Enable robust reconstruction in complex, real-world environments such as urban and off-road scenes.
• Provide a continuous, implicit surface representation suitable for terrain traversability and physical modeling tasks.
• Overcome limitations of existing methods that either oversmooth or overextend surfaces in low-density regions.

Proposed method

• Represents input Lidar point clouds via a regular voxel grid, storing per-voxel statistics: mean, covariance, and point count.
• Uses local planar surface approximation within adaptive neighborhoods to model surface geometry robustly.
• Applies an adaptive neighborhood kernel that dynamically selects the smallest neighborhood satisfying a confidence threshold based on Gaussian distribution statistics.
• Introduces a Gaussian confidence metric (Eq. 3) to assess the reliability of local surface estimates and prevent overextension in low-confidence regions.
• Constructs a Truncated Signed Distance Field (TSDF) from the confidence-weighted local planes to ensure smooth, continuous surface representation.
• Extracts the final triangular mesh via the Marching Cubes algorithm from the zero-crossing of the TSDF.

Experimental results

Research questions

• RQ1 How can local surface approximations be made robust to heterogeneous point density in Lidar data?
• RQ2 To what extent does an adaptive neighborhood strategy improve reconstruction accuracy and density compared to fixed or constant neighborhoods?
• RQ3 Can a confidence-based Gaussian metric effectively prevent surface overextension and artifact generation in low-signal regions?
• RQ4 How does the proposed method compare in accuracy and density to state-of-the-art methods like IMLS on both synthetic and real-world Lidar data?
• RQ5 Can the method maintain high detail near the sensor while reducing noise and artifacts in sparse regions?

Key findings

• The proposed method achieves a symmetric Hausdorff distance (AEsym) of 0.14m on the CARLA dataset, significantly outperforming IMLS (0.23m).
• At a 30m distance, the average reconstruction error of the proposed method is ≈0.2m, compared to ≈0.35m for IMLS, indicating a 50% reduction in error.
• Approximately 80% of the vertices in the reconstructed mesh have an error below 0.2m, compared to only 40% for IMLS, demonstrating superior vertex-level accuracy.
• Despite producing a sparser mesh, the method generates fewer artifacts and avoids surface extension, particularly noticeable in lampposts, signs, and walls in qualitative comparisons.
• The combination of adaptive neighborhood and Gaussian confidence (AN+GC) yields the best balance: higher accuracy (lower P→GT) and better density (lower GT→P) than constant neighborhood or confidence-only approaches.
• On the KITTI dataset, the method maintains structural fidelity and avoids over-smoothing or over-extension, confirming its robustness on real-world data.