[Paper Review] GSeg3D: A High-Precision Grid-Based Algorithm for Safety-Critical Ground Segmentation in LiDAR Point Clouds

Muhammad Haider Khan Lodhi, Christoph Hertzberg|arXiv (Cornell University)|Mar 4, 2026

Remote Sensing and LiDAR Applications0 citations

TL;DR

GSeg3D introduces a two-phase grid-based ground segmentation method with KD-tree–driven region expansion to achieve high precision and robust recall in LiDAR point clouds, validated on SemanticKITTI. It emphasizes safety-critical performance with stable runtime.

ABSTRACT

Ground segmentation in point cloud data is the process of separating ground points from non-ground points. This task is fundamental for perception in autonomous driving and robotics, where safety and reliable operation depend on the precise detection of obstacles and navigable surfaces. Existing methods often fall short of the high precision required in safety-critical environments, leading to false detections that can compromise decision-making. In this work, we present a ground segmentation approach designed to deliver consistently high precision, supporting the stringent requirements of autonomous vehicles and robotic systems operating in real-world, safety-critical scenarios.

Motivation & Objective

• Address the need for high-precision ground segmentation in LiDAR data for safety-critical autonomous systems.
• Develop a robust method that minimizes false positives while maintaining strong recall across diverse environments.
• Leverage a two-phase grid representation and local geometric analysis to improve segmentation accuracy and robustness.
• Evaluate performance against state-of-the-art methods on SemanticKITTI and analyze precision, recall, and F1-score across distances.

Proposed method

• Two-phase ground segmentation: Phase I uses a large-height grid to quickly mark non-ground structures; Phase II uses a smaller-height grid to refine tentative ground points.
• Grid-based representation assigns points to a regular 3D grid using cellsize and computes a 3D covariance per cell for local geometry.
• Local eigen classification within each grid cell (line, planar, non-planar) based on eigenvalues of the covariance matrix to guide ground estimation.
• Surface gradient analysis with RANSAC plane fitting in planar cells to estimate local slope and classify tentative ground vs non-ground.
• Ground Region Expansion uses a KD-tree of tentative ground centroids for radius-based neighborhood expansion starting from a seed cell under the robot, enabling high-resolution connectivity.
• Multi-step refinement post-expansion including inlier/outlier separation, bounding-box sparsity analysis, ambiguity resolution via local neighborhood, and floating-cell rejection to ensure spatially coherent ground regions.

Experimental results

Research questions

• RQ1 Can a dual-phase, grid-based segmentation maintain high precision while preserving recall across diverse cluttered and vegetation-rich scenes?
• RQ2 Does KD-tree–driven region expansion enable robust ground connectivity at finer grid resolutions compared to traditional neighbor expansion?
• RQ3 How does GSeg3D compare to state-of-the-art ground segmentation methods in terms of precision, recall, F1-score and runtime on SemanticKITTI?
• RQ4 What are the limitations of pure geometric methods in vegetation-heavy environments, and can refinement steps mitigate these?

Key findings

• GSeg3D achieves high precision (mean 96.6%) with low variability (std 2.7%) across all sequences.
• GSeg3D delivers strong recall (mean 89.4%) and competitive F1-score (92.8%), balancing accuracy and completeness.
• Compared to Linefit, GSeg3D offers substantially better recall while maintaining near-top precision and stable performance.
• In overall runtime, GSeg3D (48.1 ms) trades off some speed for robust segmentation quality, remaining practical for real-time applications.
• Across urban, suburban, highway, and vegetation-rich scenes, GSeg3D maintains high precision (>94%) with robust F1-scores (89–93%).
• GSeg3D’s two-phase approach and KD-tree expansion help preserve connectivity and mitigate gaps due to sparse or non-uniform LiDAR scans.