The purpose of this tutorial is to provide examples of how to work with 3D or multidimensional data using two popular libraries: Point Cloud Library (PCL) and Open3D. These examples will cover such topics as I/O, features, keypoints, registration, segmentation, and sample consensus.
Your system should have a modern compiler that supports C++11, as well as git and CMake. To use the python bindings for Open3D, you will also need either python 2.7 or 3.5+.
If you have ROS installed already, then you should have a version of PCL on your computer already.
If not, then you can install pre-built binaries (or build from source) for a variety of operating systems based on these instructions: http://www.pointclouds.org/downloads/.
The Ubuntu directions are copied below for convenience:
sudo add-apt-repository ppa:v-launchpad-jochen-sprickerhof-de/pcl
sudo apt-get update
sudo apt-get install libpcl-all
Even if you have a version of the library through ROS, you should also install some tools for interfacing with the library in a non-ROS way:
sudo apt-get install pcl-tools
This library can be installed from source, according to these directions http://www.open3d.org/docs/getting_started.html#ubuntu.
However, I found that for Ubuntu 16.04, there are conflicts between ROS and libpng16-dev, which is one of the normal dependencies if you follow the instructions.
Executing their script to install dependencies will remove ROS, which you probably don't want.
So here is a workaround that I had success with on 16.04:
git clone https://github.com/IntelVCL/Open3D
sudo pip (or pip3) install pybind11
sudo apt-get install xorg-dev libglu1-mesa-dev libgl1-mesa-glx libglew-dev libglfw3-dev libjsoncpp-dev libeigen3-dev libjpeg-dev python-dev python-tk python3-dev python3-tk
cd Open3D
mkdir build
cd build
cmake -DPYTHON_EXECUTABLE:FILEPATH=/usr/bin/python (or /usr/bin/python3) -DBUILD_PNG=YES ../src
make -j
sudo make install
You can then verify that the library installed correctly by opening a python interpreter and checking to see that import open3d does not produce errors.
- If you're reading this README, you've probably already cloned this repository, but if not, navigate to your desired install directory and run:
git clone [email protected]:jeffdelmerico/pointcloud_tutorial.git
- Navigate into the pointcloud_tutorials directory, then execute the following commands to build the examples:
mkdir build
cd build
cmake ..
make -j && make install
- Assuming the examples all built correctly, you should have some executables in a new
bindirectory.
cd ../bin
Write a pcd file of randomly generated points and visualize it with pcd_viewer:
./pcl_write_pcd
pcl_viewer test_pcd.pcd
Read in that same pcd file:
./pcl_read_pcd
Grab point clouds from an RGBD camera (be sure one is connected):
./pcl_openni_grabber
Read and then write a copy of a pcd, ply, and jpg file:
python3 open3d_file_io.py
Compute point normals on a point cloud and use built-in visualizer:
./pcl_compute_normals
Compute point normals on a point cloud and use built-in visualizer:
python3 open3d_compute_normals.py
press 'n' to visualize the normals once they have been computed.
Run one of 3 different filters on a point cloud:
./pcl_filtering 0 (pass through filter)
./pcl_filtering 1 (downsample to a voxel grid)
./pcl_filtering 2 (perform statistical outlier removal)
Visualize the output side-by-side with the original:
pcl_viewer -multiview 1 ../data/table_scene_lms400.pcd table_scene_lms400_filtered.pcd
press 'r' to zero the viewpoint, and 'l' to list the color handlers.
Run the first two filters sequentially:
python3 open3d_filtering.py
Outlier removal filtering is not supported, and it's also not straightforward to visualize multiple point clouds separately.
Find SIFT keypoints in a point cloud and visualize:
./pcl_keypoints ../data/robot1.pcd keypoints
Compute PFH features on SIFT keypoints for two point clouds and then compute their correspondences:
./pcl_keypoints ../data/robot correspondences
Open3D uses downsampled point clouds rather than keypoints when computing features and correspondences. However, it is possible to compute PFH features on a downsampled point cloud:
python3 open3d_keypoints.py
Put some random points into a KdTree; do NN and radius search near a random point in space:
./pcl_kdtree
Put some random points into an Octree; do voxel, NN, and radius search near a random point in space:
./pcl_octree
Load a point cloud into a KdTree, then do NN and radius search around a point in the cloud:
python3 open3d_kdtree.py
Generate some points that fit a planar model as well as a bunch of outliers:
./pcl_sample_consensus
Generate points as before, but use sample consensus to find inliers to a planar model:
./pcl_sample_consensus -f
Plane fitting is not implemented in Open3D, but would be straightforward to implement. RANSAC is used implicitly within Open3D's registration functionaity.
Perform iterative plane segmentation on real point cloud data:
./pcl_plane_segmentation
Visualize the output side-by-side with the original:
pcl_viewer -multiview 1 ../data/table_scene_lms400.pcd table_scene_lms400_first_plane.pcd table_scene_lms400_second_plane.pcd
Perform euclidean cluster extraction after removing the dominant planes in the scene:
./pcl_euclidean_cluster_extraction
Visualize the output with all clusters in the same viewport:
pcl_viewer cloud_cluster_0.pcd cloud_cluster_1.pcd cloud_cluster_2.pcd cloud_cluster_3.pcd cloud_cluster_4.pcd
Operations for plane segmentation and clustering are not implemented in Open3D, but the algorithms would be straightforward to implement.
Perform iterative closest point to align two point clouds:
./pcl_icp ../data/cloud_bin_0.pcd ../data/cloud_bin_1.pcd
Visualize aligned and combined point cloud beside originals:
pcl_viewer -multiview 1 ../data/pcl_icp_init.pcd ../data/pcl_icp_aligned.pcd
Attempt to fit several point cloud templates to the target point cloud, output the best match:
./pcl_template_matching ../data/object_templates.txt ../data/person.pcd
Visualize the matched and aligned template against the target PC:
pcl_viewer ../data/person.pcd template_aligned.pcd
you may need to press '1' several times to get a good color scheme for the two point clouds to be visible.
Open3D offers implementations of several algorithms for both local and global point cloud registration.
Local ICP with an initial guess can be performed with either point-to-point or point-to-plane alignment:
python3 open3d_icp.py
There is also a variant of the ICP algorithm within Open3D that utilizes color as well as geometric information:
python3 open3d_colored_icp.py
These local methods rely on a reasonably close initial guess.
For global registration, an initial coarse registration is computed using feature correspondences, in order to bootstrap a finer alignment step:
python3 open3d_global_registration.py
Finally, there is also another algorithm to compute this coarse initial registration, which is significantly faster than the standard RANSAC. This script compares the two approaches:
python3 open3d_fast_global_registration.py
Some of the PCL examples here were adapted from these tutorials: http://www.pointclouds.org/documentation/tutorials/, but there are many more topics covered there. The Open3D examples were also adapted from the official tutorials http://www.open3d.org/docs/tutorial/index.html where there are further examples.