Class COpenNI2Sensor

Defined in File COpenNI2Sensor.h

Inheritance Relationships

Base Types

public mrpt::hwdrivers::CGenericSensor (Class CGenericSensor)
public mrpt::hwdrivers::COpenNI2Generic (Class COpenNI2Generic)

Class Documentation

class COpenNI2Sensor : public mrpt::hwdrivers::CGenericSensor, public mrpt::hwdrivers::COpenNI2Generic 

A class for grabbing “range images”, intensity images (either RGB or IR) and other information from an OpenNI2 sensor. This class permits one to access several sensors simultaneously. The same options (resolution, fps, etc.) are used for every sensor.

Configuration and usage:

Data is returned as observations of type mrpt::obs::CObservation3DRangeScan. See those classes for documentation on their fields.

As with any other CGenericSensor class, the normal sequence of methods to be called is:

CGenericSensor::loadConfig() - Or calls to the individual setXXX() to configure the sensor parameters.
COpenNI2Sensor::initialize() - to start the communication with the sensor.
call COpenNI2Sensor::getNextObservation() for getting the data.

Calibration parameters

In this class we employ the OpenNI2 method to return depth images referred to the RGB camera. Otherwise we could specify an accurate transformation of depth images to 3D points, you’ll have to calibrate your RGBD sensor for that, and supply the following threee pieces of information (default calibration data will be used otherwise, but they’ll be not optimal for all sensors!):

Camera parameters for the RGB camera. See COpenNI2Sensor::setCameraParamsIntensity()
Camera parameters for the depth camera. See COpenNI2Sensor::setCameraParamsDepth()
The 3D relative pose of the two cameras. See COpenNI2Sensor::setRelativePoseIntensityWrtDepth()

See https://www.mrpt.org/Kinect_calibration for a procedure to calibrate RGBD sensors with an interactive GUI program.

Coordinates convention

The origin of coordinates is the focal point of the depth camera, with the axes oriented as in the diagram shown in mrpt::obs::CObservation3DRangeScan. Notice in that picture that the RGB camera is assumed to have axes as usual in computer vision, which differ from those for the depth camera.

The X,Y,Z axes used to report the data from accelerometers coincide with those of the depth camera (e.g. the camera standing on a table would have an ACC_Z=-9.8m/s2).

Notice however that, for consistency with stereo cameras, when loading the calibration parameters from a configuration file, the left-to-right pose increment is expected as if both RGB & IR cameras had their +Z axes pointing forward, +X to the right, +Y downwards (just like it’s the standard in stereo cameras and in computer vision literature). In other words: the pose stored in this class uses a different axes convention for the depth camera than in a stereo camera, so when a pose L2R is loaded from a calibration file it’s actually converted like:

L2R(this class convention) = CPose3D(0,0,0,-90deg,0deg,-90deg) (+) L2R(in the config file)

Some general comments

Depth is grabbed in millimeters
This sensor can be also used from within rawlog-grabber to grab datasets within a robot with more sensors.
There is no built-in threading support, so if you use this class manually (not with-in rawlog-grabber), the ideal would be to create a thread and continuously request data from that thread (see mrpt::system::createThread ).
The intensity channel default to the RGB images, but it can be changed with setVideoChannel() to read the IR camera images (useful for calibrating).
There is a built-in support for an optional preview of the data on a window, so you don’t need to even worry on creating a window to show them.
This class relies on an embedded version of libfreenect (you do NOT need to install it in your system). Thanks guys for the great job!

Converting to 3D point cloud

You can convert the 3D observation into a 3D point cloud with this piece of code:

mrpt::obs::CObservation3DRangeScan  obs3D;
mrpt::maps::CColouredPointsMap       pntsMap;
pntsMap.colorScheme.scheme = CColouredPointsMap::cmFromIntensityImage;
pntsMap.loadFromRangeScan(obs3D, std::nullopt);

Then the point cloud mrpt::maps::CColouredPointsMap can be converted into an OpenGL object for rendering with mrpt::maps::CMetricMap::getAs3DObject() or alternatively with:

     mrpt::viz::CPointCloudColoured::Ptr gl_points =
*mrpt::viz::CPointCloudColoured::Create();
     gl_points->loadFromPointsMap(&pntsMap);

Platform-specific comments

For more details, refer to libfreenect documentation:

Linux: You’ll need root privileges to access Kinect. Or, install MRPT/scripts/51-kinect.rules in /etc/udev/rules.d/ to allow access to all users.
Windows:
- Since MRPT 0.9.4 you’ll only need to install libusb-win32: download and extract the latest libusb-win32-bin-x.x.x.x.zip
- To install the drivers, read this: http://openkinect.org/wiki/Getting_Started#Windows
MacOS: (write me!)

Format of parameters for loading from a .ini file

   PARAMETERS IN THE ".INI"-LIKE CONFIGURATION STRINGS:
  -------------------------------------------------------
    [supplied_section_name]
     sensorLabel     = OPENNI2       // A text description
     preview_window  = false        // Show a window with a preview of the
*grabbed data in real-time
 
     device_number   = 0           // Device index to open (0:first Kinect,
 1:second Kinect,...)
 
     grab_image      = true        // Grab the RGB image channel?
*(Default=true)
     grab_depth      = true        // Grab the depth channel? (Default=true)
     grab_3D_points  = true        // Grab the 3D point cloud? (Default=true)
*If disabled, points can be generated later on.
 
     video_channel   = VIDEO_CHANNEL_RGB // Optional. Can be:
*VIDEO_CHANNEL_RGB (default) or VIDEO_CHANNEL_IR
 
     pose_x=0   // Camera position in the robot (meters)
     pose_y=0
     pose_z=0
     pose_yaw=0 // Angles in degrees
     pose_pitch=0
     pose_roll=0
 
 
     // Kinect sensor calibration:
     // See https://www.mrpt.org/Kinect_and_MRPT
 
     // Left/Depth camera
     [supplied_section_name_LEFT]
     rawlog-grabber-ignore = true // Instructs rawlog-grabber to ignore this
*section (it is not a separate device!)
 
     resolution = [640 488]
     cx         = 314.649173
     cy         = 240.160459
     fx         = 572.882768
     fy         = 542.739980
     dist       = [-4.747169e-03 -4.357976e-03 0.000000e+00 0.000000e+00
 0.000000e+00]    // The order is: [K1 K2 T1 T2 K3]
 
     // Right/RGB camera
     [supplied_section_name_RIGHT]
     rawlog-grabber-ignore = true // Instructs rawlog-grabber to ignore this
*section (it is not a separate device!)
 
     resolution = [640 480]
     cx         = 322.515987
     cy         = 259.055966
     fx         = 521.179233
     fy         = 493.033034
     dist       = [5.858325e-02 3.856792e-02 0.000000e+00 0.000000e+00
 0.000000e+00]    // The order is: [K1 K2 T1 T2 K3]
 
     // Relative pose of the right camera wrt to the left camera:
     // This assumes that both camera frames are such that +Z points
     // forwards, and +X and +Y to the right and downwards.
     // For the actual coordinates employed in 3D observations, see figure in
*mrpt::obs::CObservation3DRangeScan
     [supplied_section_name_LEFT2RIGHT_POSE]
     rawlog-grabber-ignore = true // Instructs rawlog-grabber to ignore this
*section (it is not a separate device!)
 
     pose_quaternion      = [0.025575 -0.000609 -0.001462 0.999987 0.002038
 0.004335 -0.001693]