CurbNet: Curb Detection Framework Based on LiDAR Point Cloud Segmentation

---

📌 Overview

Curb detection is a critical task in autonomous driving and robotic perception, enabling accurate understanding of road boundaries and drivable areas.

However, it remains challenging due to:

• (a) Difficulty in extracting height-aware geometric features
• (b) Non-uniform density distribution of LiDAR point clouds
• (c) Severe class imbalance between curb and non-curb points

To address these challenges, we propose CurbNet, a LiDAR-based curb detection framework with:

• A newly constructed 3D-Curb dataset
• A Multi-Scale Channel Attention (MSCA) module for feature fusion and height modeling
• A loss group strategy to handle class imbalance
• A post-processing refinement module for improved prediction quality

🧠 Framework

📊 Detection Results

SemanticKITTI Benchmark

🗂️ Dataset: 3D-Curb

We construct the 3D-Curb dataset based on the large-scale SemanticKITTI dataset, with the following features:

• Add a new curb category (label = 3) with full 3D annotations
• Preserve original 28 semantic classes
• Captured using 64-line LiDAR
• Designed for urban autonomous driving scenarios

📥 Download

• Baidu Netdisk: Download Link (Code: 1234)
• Google Drive: Download Link

📦 Data Format

The dataset follows the SemanticKITTI format:

/kitti/dataset/
└── sequences/
├── 00/
│   ├── labels/
│   │     ├ 000000.label
│   │     └ 000001.label
│   └── velodyne/
│         ├ 000000.bin
│         └ 000001.bin
├── 01/
├── 02/
...
└── 10/

• Point clouds: .bin
• Labels: .label

⚠️ Important Note

Since curb is an additional category, we assign:

label: 3 → "curb"

To visualize correctly using SemanticKITTI API, please modify:

config/semantic-kitti.yaml

Add:

3: "curb"

🔧 Visualization Tool

We recommend using the official SemanticKITTI API:

👉 https://github.com/PRBonn/semantic-kitti-api

📁 Additional Dataset: NRS

The original NRS dataset contains only 2D annotations.

We:

• Project 2D labels to 3D point clouds
• Convert format to SemanticKITTI style
• Release processed version for public use

📥 Download

• Baidu Netdisk: Download Link (Code: 1234)
• Google Drive: Download Link

---

⚙️ Installation

Requirements

• Python ≥ 3.7
• PyTorch ≥ 1.2
• yaml
• Cython
• torch-scatter
• spconv (tested with spconv==1.2.1, cuda==10.2)
• nuScenes-devkit (optional)

🚀 Training

1. Modify config file:

config/semantickitti-curb_0.2.yaml

2. Start training:

sh train_0.2.sh

🎯 Inference Demo

Run demo on a folder of LiDAR scans:

python demo_folder_focal.py \
    --demo-folder YOUR_FOLDER \
    --save-folder YOUR_SAVE_FOLDER

With labels (optional):

python demo_folder_focal.py \
    --demo-folder YOUR_FOLDER \
    --save-folder YOUR_SAVE_FOLDER \
    --demo-label-folder YOUR_LABEL_FOLDER

---

⭐ Citation

If you find this project useful, please consider giving a ⭐ and citing our work:

@article{zhao2025curbnet,
  title={CurbNet: Curb detection framework based on LiDAR point cloud segmentation},
  author={Zhao, Guoyang and Ma, Fulong and Qi, Weiqing and Liu, Yuxuan and Liu, Ming and Ma, Jun},
  journal={IEEE Transactions on Intelligent Transportation Systems},
  year={2025},
  publisher={IEEE}
}

📬 Contact

If you have any questions or suggestions, feel free to open an issue or contact the authors.

🌟 Acknowledgements

This project is built upon:

• SemanticKITTI dataset
• Open-source LiDAR perception frameworks Cylinder3D

We thank the community for their contributions.