Towards Robust and Efficient Ground Segmentation: A Dual-Stage Approach With Dynamic Parameter Optimization
- 1 April 2026
- journal article
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Vehicular Technology
- Vol. PP (00189545) , 1-16
Abstract
Ground segmentation from LiDAR point clouds is a critical task in autonomous driving and intelligent transportation systems, as it directly affects road surface estimation, obstacle detection, and path planning. Existing methods, however, often degrade on sloped or irregular terrains and fail to preserve clear ground–object boundaries. To address these limitations, we propose a dual-stage ground segmentation framework with dynamic parameter optimization. In the coarse stage, adaptive plane modeling with slope-aware parameter adjustment enables efficient separation of ground candidates. In the fine stage, spherical depth map restoration andmulti-feature fusion refine boundary localization and recover fine-grained details. Extensive experiments on SemanticKITTI and Rellis-3D show that our method achieves 95.90% accuracy with 14.59 ms latency per scan in structured urban roads, while sustaining 90.64% accuracy in unstructured off-road environments. These results demonstrate that the proposed approach delivers both robustness and efficiency, meeting real-time requirements for autonomous driving applications. The source code of our method will be made publicly available at: https://github.com/PKZD-JLL/DipGSegplus.Keywords
Funding Information
- Tianshan Talents Cultivation Program - Leading Talents for Scientific and Technological Innovation (2024TSYCLJ0002)
- Guangxi science and technology major project (GUIKEAA23062031)
This publication has 38 references indexed in Scilit:
- FGSeg: Field-ground segmentation for agricultural robot based on LiDARComputers and Electronics in Agriculture, 2023
- Projection-Based Point Convolution for Efficient Point Cloud SegmentationIEEE Access, 2022
- SemanticKITTI: A Dataset for Semantic Scene Understanding of LiDAR Sequences2019
- PointPillars: Fast Encoders for Object Detection From Point Clouds2019
- Frustum PointNets for 3D Object Detection from RGB-D Data2018
- VoxelNet: End-to-End Learning for Point Cloud Based 3D Object Detection2018
- Ground segmentation and free space estimation in off-road terrainPattern Recognition Letters, 2018
- DeMoN: Depth and Motion Network for Learning Monocular Stereo2017
- Efficient Online Segmentation for Sparse 3D Laser ScansPFG – Journal of Photogrammetry, Remote Sensing and Geoinformation Science, 2017
- On the segmentation of 3D LIDAR point clouds2011