PMR (point model registration) aims to register a pointcloud to a CAD model (usually described in .stl format).
This work is published in the article "Estimation of Point Cloud Object Pose Using Particle Swarm Optimization (ICMVA 2018)" https://www.researchgate.net/publication/326165587
load stl model --> downsample stl triangles --> load pcd data --> filter cloud --> compute normal --> icf estimate
the icf estimate is accelerated by PSO algorithm:
-
common -- defines data structure of stl model
-
conversion -- convert stl model to pointCloud (.pcd)
-
filter -- randomly downsample triangles of the stl model
-
io -- load stl model from file
-
math -- a set of algorithms, including pso(partical swarm optimization), and transform algorithm.
-
registration -- defines registration pipelines
- icf -- iterative closet face method
- mcn -- mean cluster normals estimation method (todo)
-
visualization -- defines stl visualizer
-
examples -- demos of icf_pso method
-
tools -- defines some utility tools
- cloud_transform -- transform a point cloud by 6D pose, including 3 euler angles(theta, phi, psi) and translation offsets(x,y,z)
./cloud_transform <stl_file_name> <0 0.5 0 10 0 0>- pcd_gen -- convert a stl model to a point cloud model, a leafsize should be given to define the point size
./pcd_gen <stl_file_name> <pcd_file_name> <leafsize>- pcd_viz -- visualize a pcd file
./pcv_viz <pcd_file_name>- stl_viz -- visualize a stl file
./stl_viz <stl_file_name>
- PCL 1.8
- VTK (v5.8 is suggested)
git clone
cd pmr
mkdir build && cd build
cmake ..
make
These commands will generate 2 demos: icf_demo and icp_demo.
icf_demo uses directly face information from stl model and points from point cloud, to determine the pose of the point cloud. PSO algorithm is applied to accelerate the search speed.
However, icp_demo firstly converts the stl model to a point cloud model, and calculates the distances of points pairs to estimate the pose.
cd tools
mkdir build && cd build
cmake ..
make
There is a testdata folder in tools, which is used for testing these tools.
icf_demo need 2 arguments: stl_model_path and pcd_file_path
cd build
./icf_demo ../models/cuboid.stl ../models/cuboid_30.pcd
The register precision and the computation time is a trade-off. If we increase the maximum iteration steps, we will ger better results, meanwhile, maybe we have to wait for the results over a cup of coffee.
After 40 steps of iteration, the cuboid point cloud was aligned with an error.
After 200 steps, the cuboid point cloud was aligned well.
If you need to adjust the maximum step of iteration, here is an instruction:
in icf_pso.cpp
void IterativeClosestFace_PSO::estimate(Eigen::Matrix4f & estimated_transf) {
math::Swarm sw;
//Config paras
math::Pose minBound(-1, -1, -1, -M_PI/2, -M_PI/2, -M_PI/2);
math::Pose maxBound(1, 1, 1, M_PI/2, M_PI/2, M_PI/2);
math::Pose minVel(-0.04, -0.04, -0.04, -M_PI/60, -M_PI/60, -M_PI/60);
math::Pose maxVel(0.04, 0.04, 0.04, M_PI/60, M_PI/60, M_PI/60);
sw.setSearchSpaceBound(minBound,maxBound);
sw.setSearchSpeedBound(minVel,maxVel);
sw.setMaxIteration(200); // increase the max step will improve the result precision
sw.setPopulationSize(20);
...
}