3D-BBS: Global Localization for 3D Point Cloud Scan Matching Using Branch-and-Bound Algorithm

This repository contains the public source code for the paper to be published at ICRA2024 in Yokohama.

If you like the 3D-BBS, we kindly ask for your support by giving our repository star ⭐!

@inproceedings{aoki20243dbbs,
  title={3D-BBS: Global Localization for 3D Point Cloud Scan Matching Using Branch-and-Bound Algorithm}, 
  author={Koki Aoki and Kenji Koide and Shuji Oishi and Masashi Yokozuka and Atsuhiko Banno and Junichi Meguro},
  booktitle={IEEE International Conference on Robotics and Automation},
  year={2024},
  organization={IEEE}
}

🔍 Quick understanding about 3D-BBS

Task

3D global localization for such as robot's initial pose estimation

Input

A single 3D LiDAR scan aligned in the gravity direction by using such as IMU
(Although 3D-BBS also performs 6DoF search without gravity aligning, the processing time is more than 10 times longer.)
3D pre-built point cloud map

Output

$4\times4$ transformation matrix about coarse sensor pose in map coordinates.

Main Features

Full search algorithm based on branch-and-bound (BnB) scan matching
Fast processing (latest: about 20 ms ~) by proposed GPU-accelerated BnB and roto-translational branching
No initial pose (we assume that the roll and pitch angles are roughly estimated by IMU for faster processing)
Low voxelmap memory consumption due to sparse hash map

Other Features

Click here!

Optional Voxelmap pre-construction and direct read function
Optional timeout function
Minimal dependencies in core source code so that user can easily integrate into the other projects
Docker support by @Taeyoung96
Test data is available, allowing evaluation of localization error and processing time.
ROS2 implementation example

📣 News

The latest implementation demonstrates faster processing time than our paper. Performance on our test data:

Hierarchical voxelmap construction
- Paper: 9,272 ms on average
- Latest: 3,494 ms on average
- Load saved voxelmaps directly: 130 ms on average (See step 5 on this page to save voxelmaps)
Global localization
- Paper: 878 ms on average
- Latest: 189 ms on average

The hardware configuration (Intel Core i7-10700K 3.8GHz, 32GB RAM, and NVIDIA GeForce RTX2060) is the same as in the paper.

✅ Dependencies

bbs3d (Lower versions are not tested)
- CMake
- Eigen3 (3.4.0 or higher)
- OpenMP
- CUDA (12.0 or higher)
test
- (All bbs3d dependencies)
- PCL
ros2_test
- (All bbs3d dependencies)
- ros2 humble

Support Docker 🐳

If nvidia driver is 525.60.11 or higher, try docker!
For more information, you can check docker_start.md

🌟 3D-BBS core source code

Build and Install

# Note: If you are using Eigen3 below 3.4.0, git clone with --recursive
git clone https://github.com/KOKIAOKI/3d_bbs.git
cd 3d_bbs
mkdir build && cd build

CPU ver. & GPU ver. (Please ignore the large number of warnings)

cmake .. -DCMAKE_BUILD_TYPE=Release
make -j
sudo make install

CPU ver. only

cmake .. -DCMAKE_BUILD_TYPE=Release -DBUILD_CUDA=OFF
make -j
sudo make install

📋 Test code

See test_code.md

🤖 ROS 2 test code

See ros2_test_code.md

💻 Use 3D-BBS in your cmake project