Reproduce “GroundGrid:LiDAR Point Cloud Ground Segmentation and Terrain Estimation” on OpenResearch

About this paper

Nicolai Steinke, Daniel Göhring, Raùl Rojas

GroundGrid, developed by Freie Universität Berlin, presents a deterministic, grid-based approach for LiDAR ground segmentation and terrain estimation that achieves state-of-the-art performance without extensive training data. It leverages explicit outlier filtering, adaptive distance-dependent variance thresholds, and temporal integration to accurately distinguish ground from non-ground and estimate terrain, achieving an IoU of 94.78% on SemanticKITTI and operating at 171 Hz.

Key insights

• Reflection-induced outlier points, common in LiDAR data, can be effectively identified and filtered using line-of-sight checks against existing elevation maps.
• Adaptive, distance-dependent variance thresholds combined with neighborhood analysis improve ground classification robustness across varying sensor ranges and noise levels.
• Confidence-weighted temporal integration of ground observations enhances the stability and long-term accuracy of terrain elevation estimates, adapting to changing environments.

Summary by alphaXiv.