fast_gicp

This package is a collection of GICP-based fast point cloud registration algorithms. It constains a multi-threaded GICP as well as multi-thread and GPU implementations of our voxelized GICP (VGICP) algorithm. All the implemented algorithms have the PCL registration interface so that they can be used as an inplace replacement for GICP in PCL.

FastGICP: multi-threaded GICP algorithm (~40FPS)
FastGICPSingleThread: GICP algorithm optimized for single-threading (~15FPS)
FastVGICP: multi-threaded and voxelized GICP algorithm (~70FPS)
FastVGICPCuda: CUDA-optimized voxelized GICP algorithm (~120FPS)

on melodic & noetic

Installation

Dependencies

PCL
Eigen
OpenMP
CUDA (optional)
Sophus
nvbio

We have tested this package on Ubuntu 18.04/20.04 and CUDA 11.1.

CUDA

To enable the CUDA-powered implementations, set BUILD_VGICP_CUDA cmake option to ON.

ROS

cd ~/catkin_ws/src
git clone https://github.com/SMRT-AIST/fast_gicp --recursive
cd .. && catkin_make -DCMAKE_BUILD_TYPE=Release
# enable cuda-based implementations
# cd .. && catkin_make -DCMAKE_BUILD_TYPE=Release -DBUILD_VGICP_CUDA=ON

Non-ROS

git clone https://github.com/SMRT-AIST/fast_gicp --recursive
mkdir fast_gicp/build && fast_gicp/build
cmake .. -DCMAKE_BUILD_TYPE=Release
# enable cuda-based implementations
# cmake .. -DCMAKE_BUILD_TYPE=Release -DBUILD_VGICP_CUDA=ON
make -j8

Benchmark

CPU:Core i9-9900K GPU:GeForce RTX2080Ti

roscd fast_gicp/data
rosrun fast_gicp gicp_align 251370668.pcd 251371071.pcd

target:17249[pts] source:17518[pts]
--- pcl_gicp ---
single:127.508[msec] 100times:12549.4[msec] fitness_score:0.204892
--- pcl_ndt ---
single:53.5904[msec] 100times:5467.16[msec] fitness_score:0.229616
--- fgicp_st ---
single:111.324[msec] 100times:10662.7[msec] 100times_reuse:6794.59[msec] fitness_score:0.204379
--- fgicp_mt ---
single:20.1602[msec] 100times:1585[msec] 100times_reuse:1017.74[msec] fitness_score:0.204412
--- vgicp_st ---
single:112.001[msec] 100times:7959.9[msec] 100times_reuse:4408.22[msec] fitness_score:0.204067
--- vgicp_mt ---
single:18.1106[msec] 100times:1381[msec] 100times_reuse:806.53[msec] fitness_score:0.204067
--- vgicp_cuda (parallel_kdtree) ---
single:15.9587[msec] 100times:1451.85[msec] 100times_reuse:695.48[msec] fitness_score:0.204061
--- vgicp_cuda (gpu_bruteforce) ---
single:53.9113[msec] 100times:3463.5[msec] 100times_reuse:1703.41[msec] fitness_score:0.204049
--- vgicp_cuda (gpu_rbf_kernel) ---
single:5.91508[msec] 100times:590.725[msec] 100times_reuse:226.787[msec] fitness_score:0.20557

See src/align.cpp for the detailed usage.

Test on KITTI

# Perform frame-by-frame registration
rosrun fast_gicp gicp_kitti /your/kitti/path/sequences/00/velodyne

Related packages

Papers

Kenji Koide, Masashi Yokozuka, Shuji Oishi, and Atsuhiko Banno, Voxelized GICP for fast and accurate 3D point cloud registration [link]

Contact

Kenji Koide, k.koide@aist.go.jp

Robot Innovation Research Center, National Institute of Advanced Industrial Science and Technology, Japan [URL]

hanzheteng/fast_gicp