ground_segmentation

GSeg3D

ROS 2   •   Paper

---

High-Precision Grid-Based Ground Segmentation for Safety-Critical Autonomous Driving and Robotics Applications

This module implements the core algorithmic components of GSeg3D, a high-precision, grid-based ground segmentation method for LiDAR point clouds designed for safety-critical autonomous driving and robotics applications.

If you use this work, please cite:

@inproceedings{lodhi2025gseg3d,
  author    = {Muhammad Haider Khan Lodhi and Christoph Hertzberg},
  title     = {GSeg3D: A High-Precision Grid-Based Algorithm for Safety-Critical Ground Segmentation in LiDAR Point Clouds},
  booktitle = {2025 7th International Conference on Robotics and Computer Vision (ICRCV)},
  pages     = {119-126},
  year      = {2025},
  doi       = {10.1109/ICRCV67407.2025.11349133}
}

Please find below a summary comparison of the average results on the SemanticKITTI sequences 00–10:

Motivation

Reliable ground segmentation is a fundamental prerequisite for:

• Obstacle detection
• Traversability estimation
• Navigation and planning
• Mapping and localization

False positives (classifying ground as obstacles) can directly compromise safety and downstream decision-making.
GSeg3D is explicitly designed to maximize precision while maintaining robust recall, even in cluttered and unstructured environments.

Algorithm Overview

GSeg3D paper performs two-phase grid-based ground segmentation. For a reference implementation of the two-phase approach, see the ROS 2 node available here.

<a href="https://github.com/dfki-ric/ground_segmentation_ros2/blob/c370b2be20c75dc9fb9e6710f933ab1328fa0981/src/ground_segmentation_ros2_node.cpp#L84C8-L84C29" >Phase I – Coarse Segmentation</a>

• Uses larger vertical grid cells
• Aggressively captures elevated structures as non-ground
• Ensures high initial precision
• May temporarily over-segment ground in cluttered areas

<a href="https://github.com/dfki-ric/ground_segmentation_ros2/blob/c370b2be20c75dc9fb9e6710f933ab1328fa0981/src/ground_segmentation_ros2_node.cpp#L91C8-L91C30" >Phase II – Refinement</a>

• Uses smaller vertical grid cells
• Re-evaluates ground points from Phase I
• Corrects false positives and false negatives from Phase I
• Enforces vertical consistency constraints

This dual-phase strategy strikes a strong balance between precision and recall, a trade-off that is especially important in safety-critical systems.

Processing Pipeline

Each phase follows the same four core steps:

1. Grid Representation
   • The point cloud is discretized into a regular 3D grid
   • Each point is assigned to a cell based on configurable cell sizes
2. Local Eigen Classification
   • Per-cell covariance matrix and eigen decomposition
   • Cells classified as:
     • LINE (dominant linear structure)
     • PLANE (surface-like structure)
     • NOISE (scattered points)
3. Surface Gradient Analysis
   • Robust plane fitting using PCL SACSegmentation (PROSAC)
   • Slope estimation relative to gravity (world frame)
   • Planar cells exceeding slope thresholds are rejected as ground
4. Ground Region Expansion
   • Candidate ground cell centroids are indexed using a KD-tree (nanoflann)
   • Radius-based neighborhood expansion ensures connectivity
   • Works even when grid resolution is finer than LiDAR scan-line spacing
   • Phase II additionally enforces height-difference constraints

Key Design Contributions

KD-Tree–Based Ground Expansion

Traditional grid-neighbor expansion fails at high grid resolutions due to scan-line gaps.
GSeg3D overcomes this by:

• Indexing ground centroids in a KD-tree
• Performing radius-based spatial expansion
• Enabling fine grid resolutions without loss of connectivity or recall

Robust Seed Initialization

• Synthetic seed points are injected directly beneath the robot
• Ensures reliable ground initialization even with occlusions or sparse data
• Synthetic points are removed before final output and evaluation

Multi-Step Ground Verification

Each candidate cell undergoes:

• Plane inlier / outlier separation
• Bounding-box sparsity analysis
• Neighborhood height consistency checks
• Floating-cell rejection (non-ground below)

Only physically plausible and spatially coherent cells are labeled as ground.

Implementation Details

To generate the HTML documentation for this library:

System Requirements

OS: Ubuntu 22.04, Ubuntu 24.04

Dependencies

• CMake
• PCL
  
• Eigen3
  
• GoogleTest
  
• NanoFlann

Example (Ubuntu):

sudo apt update
sudo apt install cmake libpcl-dev libeigen3-dev libgtest-dev libnanoflann-dev

Build Instructions

git clone https://github.com/dfki-ric/ground_segmentation.git
cd ground_segmentation && mkdir -p build
cd build
cmake .. -DCMAKE_BUILD_TYPE=RELEASE
make

Generating Documentation

1. Make sure you have Doxygen and Graphviz installed:

sudo apt install doxygen graphviz

1. From the root of the project, run CMake and build the doc target:

cmake --build . --target doc

1. After building, open the generated HTML documentation doc/html/index.html in your browser

‍The main page will include the README and automatically list the header files and classes.

Running Unit Tests

ctest --test-dir test/ --output-on-failure

or

./test/test_pointcloud_grid

Core Class

template<typename PointT>
class PointCloudGrid;

Main API

PointCloudGrid(const GridConfig& config);
 
void setInputCloud(
    pcl::PointCloud<PointT>::Ptr input,
    const Eigen::Quaterniond& R_body2World
);
 
std::pair<
    pcl::PointCloud<PointT>::Ptr,
    pcl::PointCloud<PointT>::Ptr
> segmentPoints();

Typical Configuration

Below is a recommended baseline setup for most outdoor mobile robotics use cases (e.g., LiDAR-based UGVs). Adjust values based on sensor height, terrain roughness, and map resolution.

Phase-I

cellSizeX: 1.0
cellSizeY: 1.0
cellSizeZ: 10.0   # Phase I
slopeThresholdDegrees: 20.0
groundInlierThreshold: 0.125
centroidSearchRadius: 5.0
distToGround: -1.723 # (e.g. SemanticKITTI)
maxGroundHeightDeviation: 0.1

Phase-II

After coarse ground detection of Phase-I, Phase II increases vertical resolution to refine classification. All parameters are kept same as Phase-I expect for the cellSizeZ

cellSizeZ: 1.0

Performance Summary (SemanticKITTI)

• Precision: ~96–99% (low variance)
• Recall: ~85–91%
• F1-score: ~92–94%
• Runtime: ~48 ms (CPU, single scan average)

GSeg3D consistently demonstrates stable, high-precision performance across:

• Urban environments
• Highways
• Unstructured terrain

Intended Use

• Safety-critical autonomous driving
• Outdoor mobile robotics
• Traversability analysis
• Mapping and perception pipelines

Notes & Future Work

• Recall degradation in dense vegetation remains challenging
• Planned extensions:
  • Semantic-aware refinement
  • Temporal fusion
  • GPU acceleration

Tests & Visualisation

This directory contains unit tests and a visual debugging tool for the PointCloudGrid ground segmentation implementation.

The code uses PCL, Eigen, and GoogleTest and is intended for quick validation of grid indexing, ground classification logic, and phase-based segmentation behaviour.

Contents

• test_pointcloud_grid.cpp
  GoogleTest-based unit tests covering:
  • Grid indexing and hashing
  • Flat ground detection
  • Vertical structures as obstacles
  • Phase 1 vs Phase 2 behaviour
  • Orientation / gravity influence
  • Robustness against empty inputs
• test_visual_segmentation.cpp
  A visual inspection tool that:
  • Loads a .pcd or .ply point cloud
  • Runs phase 1 + phase 2 ground segmentation
  • Displays ground (green) and obstacles (red) using PCLVisualizer

Notes

• Phase 1 allows vertical and slanted propagation from the robot seed cell.
• Phase 2 refines ground classification using only previously detected ground.
• Orientation (Eigen::Quaterniond) is interpreted as sensor → world rotation.

License

BSD-3 Clause License.

© DFKI Robotics Innovation Center