[Documentation] [TitleIndex] [WordIndex

A comprehensive list of PCL tutorials can be found on PCL's external website. Here are a few of the tutorials that you might want to check out:

More information about PCL integration in ROS (e.g. point cloud data types, publishing/subscribing) can be found in the ROS/PCL overview.

The following tutorial describes how to use any of the existing tutorials on http://pointclouds.org in a ROS ecosystem using nodes or nodelets.

  Show EOL distros: 

How to use a PCL tutorial in ROS

Create a ROS package

$ roscreate-pkg my_pcl_tutorial pcl pcl_ros roscpp sensor_msgs

$ catkin_create_pkg my_pcl_tutorial pcl pcl_ros roscpp sensor_msgs

This will create a new ROS package with the necessary dependencies.

Create the code skeleton

Create an empty file called src/example.cpp and paste the following code in it:

   1 #include <ros/ros.h>
   2 #include <sensor_msgs/PointCloud2.h>
   3 // PCL specific includes
   4 #include <pcl/ros/conversions.h>
   5 #include <pcl/point_cloud.h>
   6 #include <pcl/point_types.h>
   7 
   8 ros::Publisher pub;
   9 
  10 void 
  11 cloud_cb (const sensor_msgs::PointCloud2ConstPtr& input)
  12 {
  13   // ... do data processing
  14 
  15   sensor_msgs::PointCloud2 output;
  16   // Publish the data
  17   pub.publish (output);
  18 }
  19 
  20 int
  21 main (int argc, char** argv)
  22 {
  23   // Initialize ROS
  24   ros::init (argc, argv, "my_pcl_tutorial");
  25   ros::NodeHandle nh;
  26 
  27   // Create a ROS subscriber for the input point cloud
  28   ros::Subscriber sub = nh.subscribe ("input", 1, cloud_cb);
  29 
  30   // Create a ROS publisher for the output point cloud
  31   pub = nh.advertise<sensor_msgs::PointCloud2> ("output", 1);
  32 
  33   // Spin
  34   ros::spin ();
  35 }

The code above does nothing but initialize ROS, create a subscriber and a publisher for PointCloud2 data.

Add the source file to CMakeLists.txt

Edit the CMakeLists.txt file in your newly created package and add:

rosbuild_add_executable (example src/example.cpp)

add_executable(example src/example.cpp)
target_link_libraries(example ${catkin_LIBRARIES})

Download the source code from the PCL tutorial

All PCL versions prior to 2.0, have two different ways of representing point cloud data:

  • through a sensor_msgs/PointCloud2 ROS message

  • through a pcl/PointCloud<T> data structure

The following two code examples will discuss both formats.

sensor_msgs/PointCloud2

If you'd like to save yourself some copying and pasting, you can download the source file for this example here.

The sensor_msgs/PointCloud2 format was designed as a ROS message, and is the preferred choice for ROS applications. In the following example, we downsample a PointCloud2 structure using a 3D grid, thus reducing the number of points in the input dataset considerably.

To add this capability to the code skeleton above, perform the following steps:

  • visit http://www.pointclouds.org/documentation/, click on Tutorials, then navigate to the Downsampling a PointCloud using a VoxelGrid filter tutorial (http://www.pointclouds.org/documentation/tutorials/voxel_grid.php)

  • read the code and the explanations provided there. You will notice that the code breaks down essentially in 3 parts:
    • load the cloud (lines 9-19)
    • process the cloud (lines 20-24)
    • save the output (lines 25-32)
  • since we use ROS subscribers and publishers in our code snippet above, we can ignore the loading and saving of point cloud data using the PCD format. Thus, the only relevant part in the tutorial remains lines 20-24 that create the PCL object, pass the input data, and perform the actual computation:

   1   // Create the filtering object
   2   pcl::VoxelGrid<sensor_msgs::PointCloud2> sor;
   3   sor.setInputCloud (cloud);
   4   sor.setLeafSize (0.01, 0.01, 0.01);
   5   sor.filter (*cloud_filtered);

  • In these lines, the input dataset is named cloud, and the output dataset is called cloud_filtered. Copy these lines, in the code snippet above, by modifying the callback function as follows:

   1 #include <pcl/filters/voxel_grid.h>
   2 
   3 ...
   4 
   5 void 
   6 cloud_cb (const sensor_msgs::PointCloud2ConstPtr& cloud)
   7 {
   8   sensor_msgs::PointCloud2 cloud_filtered;
   9 
  10   // Perform the actual filtering
  11   pcl::VoxelGrid<sensor_msgs::PointCloud2> sor;
  12   sor.setInputCloud (cloud);
  13   sor.setLeafSize (0.01, 0.01, 0.01);
  14   sor.filter (cloud_filtered);
  15 
  16   // Publish the data
  17   pub.publish (cloud_filtered);
  18 }

Note

Since different tutorials will often use different variable names for their inputs and outputs, remember that you may need to modify the code slightly when integrating the tutorial code into your own ROS node. In this case, notice that we had to change the variable name input to cloud, and output to cloud_filtered in order to match up with the code from the tutorial we copied.

Save the output file then build:

$ rosmake my_pcl_tutorial

$ cd %TOP_DIR_YOUR_CATKIN_HOME%
$ catkin_make

Then run:

$ rosrun my_pcl_tutorial example input:=/narrow_stereo_textured/points2

or, if you're running an OpenNI-compatible depth sensor, try:

$ rosrun my_pcl_tutorial example input:=/camera/depth/points

pcl/PointCloud<T>

As with the previous example, if you want to skip a few steps, you can download the source file for this example here.

The pcl/PointCloud<T> format represents the internal PCL point cloud format. For modularity and efficiency reasons, the format is templated on the point type, and PCL provides a list of templated common types which are SSE aligned. In the following example, we estimate the planar coefficients of the largest plane found in a scene.

To add this capability to the code skeleton above, perform the following steps:

  • visit http://www.pointclouds.org/documentation/, click on Tutorials, then navigate to the Planar model segmentation tutorial (http://www.pointclouds.org/documentation/tutorials/planar_segmentation.php)

  • read the code and the explanations provided there. You will notice that the code breaks down essentially in 3 parts:
    • create a cloud and populate it with values (lines 12-30)
    • process the cloud (38-56)
    • write the coefficients (58-68)
  • since we use ROS subscribers in our code snippet above, we can ignore the first step, and just process the cloud received on the callback directly. Thus, the only relevant part in the tutorial remains lines 38-56 that create the PCL object, pass the input data, and perform the actual computation:

   1   pcl::ModelCoefficients coefficients;
   2   pcl::PointIndices inliers;
   3   // Create the segmentation object
   4   pcl::SACSegmentation<pcl::PointXYZ> seg;
   5   // Optional
   6   seg.setOptimizeCoefficients (true);
   7   // Mandatory
   8   seg.setModelType (pcl::SACMODEL_PLANE);
   9   seg.setMethodType (pcl::SAC_RANSAC);
  10   seg.setDistanceThreshold (0.01);
  11 
  12   seg.setInputCloud (cloud.makeShared ());
  13   seg.segment (inliers, coefficients);

  • In these lines, the input dataset is named cloud and is of type pcl::PointCloud<pcl::PointXYZ>, and the output is represented by a set of point indices that contain the plane together with the plane coefficients. cloud.makeShared() creates a boost shared_ptr object for the object cloud (see the pcl::PointCloud API documentation).

Copy these lines, in the code snippet above, by modifying the callback function as follows:

   1 #include <pcl/sample_consensus/model_types.h>
   2 #include <pcl/sample_consensus/method_types.h>
   3 #include <pcl/segmentation/sac_segmentation.h>
   4 
   5 ...
   6 
   7 void 
   8 cloud_cb (const sensor_msgs::PointCloud2ConstPtr& input)
   9 {
  10   // Convert the sensor_msgs/PointCloud2 data to pcl/PointCloud
  11   pcl::PointCloud<pcl::PointXYZ> cloud;
  12   pcl::fromROSMsg (*input, cloud);
  13 
  14   pcl::ModelCoefficients coefficients;
  15   pcl::PointIndices inliers;
  16   // Create the segmentation object
  17   pcl::SACSegmentation<pcl::PointXYZ> seg;
  18   // Optional
  19   seg.setOptimizeCoefficients (true);
  20   // Mandatory
  21   seg.setModelType (pcl::SACMODEL_PLANE);
  22   seg.setMethodType (pcl::SAC_RANSAC);
  23   seg.setDistanceThreshold (0.01);
  24 
  25   seg.setInputCloud (cloud.makeShared ());
  26   seg.segment (inliers, coefficients);
  27 
  28   // Publish the model coefficients
  29   pub.publish (coefficients);
  30 }

  • Notice that we added a conversion step from sensor_msgs/PointCloud2 to pcl/PointCloud<T> on the first two lines, and we also changed the variable that we publish from output to coefficients.

In addition, since we're now publishing the planar model coefficients found rather than point cloud data, we have to change our publisher type from:

  // Create a ROS publisher for the output point cloud
  pub = nh.advertise<sensor_msgs::PointCloud2> ("output", 1);

to:

  // Create a ROS publisher for the output model coefficients
  pub = nh.advertise<pcl::ModelCoefficients> ("output", 1);

Save the output file, then compile and run the code above:

$ rosrun my_pcl_tutorial example input:=/narrow_stereo_textured/points2

or, if you're running an OpenNI-compatible depth sensor, try:

$ rosrun my_pcl_tutorial example input:=/camera/depth/points

How to use a PCL tutorial in ROS

Create a ROS package

$ roscreate-pkg my_pcl_tutorial pcl_conversions pcl_ros roscpp sensor_msgs

$ catkin_create_pkg my_pcl_tutorial pcl_conversions pcl_ros roscpp sensor_msgs 

Then, modify the package.xml to add

  <build_depend>libpcl-all-dev</build_depend>
  <exec_depend>libpcl-all</exec_depend>

Create the code skeleton

Create an empty file called src/example.cpp and paste the following code in it:

   1 #include <ros/ros.h>
   2 // PCL specific includes
   3 #include <sensor_msgs/PointCloud2.h>
   4 #include <pcl_conversions/pcl_conversions.h>
   5 #include <pcl/point_cloud.h>
   6 #include <pcl/point_types.h>
   7 
   8 ros::Publisher pub;
   9 
  10 void 
  11 cloud_cb (const sensor_msgs::PointCloud2ConstPtr& input)
  12 {
  13   // Create a container for the data.
  14   sensor_msgs::PointCloud2 output;
  15 
  16   // Do data processing here...
  17   output = *input;
  18 
  19   // Publish the data.
  20   pub.publish (output);
  21 }
  22 
  23 int
  24 main (int argc, char** argv)
  25 {
  26   // Initialize ROS
  27   ros::init (argc, argv, "my_pcl_tutorial");
  28   ros::NodeHandle nh;
  29 
  30   // Create a ROS subscriber for the input point cloud
  31   ros::Subscriber sub = nh.subscribe ("input", 1, cloud_cb);
  32 
  33   // Create a ROS publisher for the output point cloud
  34   pub = nh.advertise<sensor_msgs::PointCloud2> ("output", 1);
  35 
  36   // Spin
  37   ros::spin ();
  38 }

The code above does nothing but initialize ROS, create a subscriber and a publisher for PointCloud2 data.

Add the source file to CMakeLists.txt

Edit the CMakeLists.txt file in your newly created package and add:

rosbuild_add_executable (example src/example.cpp)

add_executable(example src/example.cpp)
target_link_libraries(example ${catkin_LIBRARIES})

Download the source code from the PCL tutorial

PCL has about four different ways of representing point cloud data, so it can get a bit confusing, but we'll try to keep it simple for you. The types are:

  • sensor_msgs::PointCloud — ROS message (deprecated)

  • sensor_msgs::PointCloud2 — ROS message

  • pcl::PCLPointCloud2 — PCL data structure mostly for compatibility with ROS (I think)

  • pcl::PointCloud<T> — standard PCL data structure

In the following code examples we will focus on the ROS message (sensor_msgs::PointCloud2) and the standard PCL data structure (pcl::PointCloud<T>). However, you should also note that pcl::PCLPointCloud2 is an important and useful type as well: you can directly subscribe to nodes using that type and it will be automatically serialized to/from the sensor_msgs type. See this example to try PCLPointCloud2 yourself.

sensor_msgs/PointCloud2

If you'd like to save yourself some copying and pasting, you can download the source file for this example here. Just remember to rename the file to example.cpp or edit your CMakeLists.txt to match.

The sensor_msgs/PointCloud2 format was designed as a ROS message, and is the preferred choice for ROS applications. In the following example, we downsample a PointCloud2 structure using a 3D grid, thus reducing the number of points in the input dataset considerably.

To add this capability to the code skeleton above, perform the following steps:

  • visit http://www.pointclouds.org/documentation/, click on Tutorials, then navigate to the Downsampling a PointCloud using a VoxelGrid filter tutorial (http://www.pointclouds.org/documentation/tutorials/voxel_grid.php)

  • read the code and the explanations provided there. You will notice that the code breaks down essentially in 3 parts:
    • load the cloud (lines 9-19)
    • process the cloud (lines 20-24)
    • save the output (lines 25-32)
  • since we use ROS subscribers and publishers in our code snippet above, we can ignore the loading and saving of point cloud data using the PCD format. Thus, the only relevant part in the tutorial remains lines 20-24 that create the PCL object, pass the input data, and perform the actual computation:

   1   // Create the filtering object
   2   pcl::VoxelGrid<pcl::PCLPointCloud2> sor;
   3   sor.setInputCloud (cloud);
   4   sor.setLeafSize (0.01, 0.01, 0.01);
   5   sor.filter (*cloud_filtered);

  • In these lines, the input dataset is named cloud, and the output dataset is called cloud_filtered. We can copy this work, but remember from earlier that we said we wanted to work with the sensor_msgs class, not the pcl class. In order to do this, we're going to have to do a little bit of extra work to convert the ROS message to the PCL type. Modify the callback function as follows:

   1 #include <pcl/filters/voxel_grid.h>
   2 
   3 ...
   4 
   5 void 
   6 cloud_cb (const sensor_msgs::PointCloud2ConstPtr& cloud_msg)
   7 {
   8   // Container for original & filtered data
   9   pcl::PCLPointCloud2* cloud = new pcl::PCLPointCloud2; 
  10   pcl::PCLPointCloud2ConstPtr cloudPtr(cloud);
  11   pcl::PCLPointCloud2 cloud_filtered;
  12 
  13   // Convert to PCL data type
  14   pcl_conversions::toPCL(*cloud_msg, *cloud);
  15 
  16   // Perform the actual filtering
  17   pcl::VoxelGrid<pcl::PCLPointCloud2> sor;
  18   sor.setInputCloud (cloudPtr);
  19   sor.setLeafSize (0.1, 0.1, 0.1);
  20   sor.filter (cloud_filtered);
  21 
  22   // Convert to ROS data type
  23   sensor_msgs::PointCloud2 output;
  24   pcl_conversions::fromPCL(cloud_filtered, output);
  25 
  26   // Publish the data
  27   pub.publish (output);
  28 }

Note

Since different tutorials will often use different variable names for their inputs and outputs, remember that you may need to modify the code slightly when integrating the tutorial code into your own ROS node. In this case, notice that we had to change the variable name input to cloud, and output to cloud_filtered in order to match up with the code from the tutorial we copied.

Note that there is a slight inefficiency here. The fromPCL can be replaced with moveFromPCL to prevent copying the entire (filtered) point cloud. However, the toPCL call cannot be optimized in this way since the original input is const.

Save the output file then build:

$ rosmake my_pcl_tutorial

$ cd %TOP_DIR_YOUR_CATKIN_HOME%
$ catkin_make

Then run:

$ rosrun my_pcl_tutorial example input:=/narrow_stereo_textured/points2

or, if you're running an OpenNI-compatible depth sensor, try:

$ roslaunch openni_launch openni.launch
$ rosrun my_pcl_tutorial example input:=/camera/depth/points

You can visualize the result by running RViz:

$ rosrun rviz rviz

and adding a "PointCloud2" display. Select camera_depth_frame for the Fixed Frame (or whatever frame is appropriate for your sensor) and select output for the PointCloud2 topic. You should see a highly downsampled point cloud. For comparison, you can view the /camera/depth/points topic and see how much it has been downsampled.

pcl/PointCloud<T>

As with the previous example, if you want to skip a few steps, you can download the source file for this example here.

The pcl/PointCloud<T> format represents the internal PCL point cloud format. For modularity and efficiency reasons, the format is templated on the point type, and PCL provides a list of templated common types which are SSE aligned. In the following example, we estimate the planar coefficients of the largest plane found in a scene.

To add this capability to the code skeleton above, perform the following steps:

  • visit http://www.pointclouds.org/documentation/, click on Tutorials, then navigate to the Planar model segmentation tutorial (http://www.pointclouds.org/documentation/tutorials/planar_segmentation.php)

  • read the code and the explanations provided there. You will notice that the code breaks down essentially in 3 parts:
    • create a cloud and populate it with values (lines 12-30)
    • process the cloud (38-56)
    • write the coefficients (58-68)
  • since we use ROS subscribers in our code snippet above, we can ignore the first step, and just process the cloud received on the callback directly. Thus, the only relevant part in the tutorial remains lines 38-56 that create the PCL object, pass the input data, and perform the actual computation:

   1   pcl::ModelCoefficients coefficients;
   2   pcl::PointIndices inliers;
   3   // Create the segmentation object
   4   pcl::SACSegmentation<pcl::PointXYZ> seg;
   5   // Optional
   6   seg.setOptimizeCoefficients (true);
   7   // Mandatory
   8   seg.setModelType (pcl::SACMODEL_PLANE);
   9   seg.setMethodType (pcl::SAC_RANSAC);
  10   seg.setDistanceThreshold (0.01);
  11 
  12   seg.setInputCloud (cloud.makeShared ());
  13   seg.segment (inliers, coefficients);

  • In these lines, the input dataset is named cloud and is of type pcl::PointCloud<pcl::PointXYZ>, and the output is represented by a set of point indices that contain the plane together with the plane coefficients. cloud.makeShared() creates a boost shared_ptr object for the object cloud (see the pcl::PointCloud API documentation).

Copy these lines, in the code snippet above, by modifying the callback function as follows:

   1 #include <pcl/sample_consensus/model_types.h>
   2 #include <pcl/sample_consensus/method_types.h>
   3 #include <pcl/segmentation/sac_segmentation.h>
   4 
   5 ...
   6 
   7 void 
   8 cloud_cb (const sensor_msgs::PointCloud2ConstPtr& input)
   9 {
  10   // Convert the sensor_msgs/PointCloud2 data to pcl/PointCloud
  11   pcl::PointCloud<pcl::PointXYZ> cloud;
  12   pcl::fromROSMsg (*input, cloud);
  13 
  14   pcl::ModelCoefficients coefficients;
  15   pcl::PointIndices inliers;
  16   // Create the segmentation object
  17   pcl::SACSegmentation<pcl::PointXYZ> seg;
  18   // Optional
  19   seg.setOptimizeCoefficients (true);
  20   // Mandatory
  21   seg.setModelType (pcl::SACMODEL_PLANE);
  22   seg.setMethodType (pcl::SAC_RANSAC);
  23   seg.setDistanceThreshold (0.01);
  24 
  25   seg.setInputCloud (cloud.makeShared ());
  26   seg.segment (inliers, coefficients);
  27 
  28   // Publish the model coefficients
  29   pcl_msgs::ModelCoefficients ros_coefficients;
  30   pcl_conversions::fromPCL(coefficients, ros_coefficients);
  31   pub.publish (ros_coefficients);
  32 }

Notice that we added two conversion steps: from sensor_msgs/PointCloud2 to pcl/PointCloud<T> and from pcl::ModelCoefficients to pcl_msgs::ModelCoefficients. We also changed the variable that we publish from output to coefficients.

In addition, since we're now publishing the planar model coefficients found rather than point cloud data, we have to change our publisher type from:

  // Create a ROS publisher for the output point cloud
  pub = nh.advertise<sensor_msgs::PointCloud2> ("output", 1);

to:

  // Create a ROS publisher for the output model coefficients
  pub = nh.advertise<pcl_msgs::ModelCoefficients> ("output", 1);

Save the output file, then compile and run the code above:

$ rosrun my_pcl_tutorial example input:=/narrow_stereo_textured/points2

or, if you're running an OpenNI-compatible depth sensor, try:

$ rosrun my_pcl_tutorial example input:=/camera/depth/points

See the output with

$ rostopic echo output

2024-11-30 14:49