Fast Grid-Based Ground Segmentation of LiDAR Point Clouds Using Dual-Seed Expansion

Area of Science:

• Robotics and Autonomous Systems
• Computer Vision
• Geospatial Data Processing

Background:

• Ground segmentation is critical for LiDAR-based perception in autonomous vehicles, impacting object detection and clustering.
• Existing methods may struggle with sloped or sparse terrain, leading to misclassifications.
• Efficient and accurate ground segmentation is essential for safe and reliable autonomous driving.

Purpose of the Study:

• To propose a fast and accurate grid-based ground segmentation method for 3D LiDAR point clouds.
• To enhance the robustness of ground segmentation in challenging environments like sloped or sparse regions.
• To provide a computationally efficient preprocessing module for real-time autonomous driving perception.

Main Methods:

• Projecting 3D LiDAR point clouds onto a 2D Cartesian grid and computing cell-wise height statistics.
• Employing a dual-seed expansion approach, combining global and LiDAR-mounting-height seeds for initial ground candidate selection.
• Implementing a local height-band criterion for point-level refinement to reduce misclassifications.

Main Results:

• Achieved an average accuracy of 96.34%, F1-score of 97.11%, and ground Intersection over Union (IoU) of 94.40% on the SemanticKITTI dataset.
• Demonstrated a fast average processing time of 12.15 ms per frame across SemanticKITTI sequences 00-10.
• The method effectively handles sloped and sparse regions, reducing missed ground candidates.

Conclusions:

• The proposed dual-seed expansion grid-based method offers a highly accurate and efficient solution for LiDAR ground segmentation.
• Its speed and accuracy make it suitable for real-time applications in autonomous driving perception systems.
• The method addresses limitations of previous approaches, particularly in challenging terrain conditions.