1.22. Pointing Smart Control¶

1.22.4.1. Skeleton Merging¶

The Skeleton Merging component is responsible for transforming the tracked postures delivered by a Kinect (or other input source) into the root coordinate system and merge the inputs of multiple sources into a single output at a constant frame rate. Usually the component will check the BaseCubeOne registry for applicable input sources and derive their transformations and scopes from there. Therefore, you have to make sure to register all input sources appropriately (See Setting up a Kinect node). Moreover, it can provide a linear stabilization for noisy inputs. This component can be configured using the following additional parameters:

• --scope-base-raw-postures defines the scope on which the listener is initialized. Keep in mind that the scopes of the inputs should be subscopes of this scope. The inputs will be filtered by type also.
• --sm-frame-rate in Hz at which output is generated. The default value 30 Hz is the average frame rate of Kinect 2s.
• --sm-stabilization-factor is the percentage of the new frame that will be made up by the previous frame (for postures that occur for both). The rest is the merged result of each input source’s last frame. If the stabilization-factor is very high a delay will be noticable in the movements of the tracked postures (for 1.0 there will be no changes after the first frame) which is a side effect of the smoothing.
• --sm-registry-transformers can be specified to restrict the input sources to a specific set. The parameter has to be provided with a list of unit ids that belong to input sources registered in the BaseCubeOne registry. If it is not specified, all suitable devices configs of the device classes specified in the next parameter will be considered as input sources.
• --sm-device-classes is a list of unit ids of device classes whose instances can be used as input sources. The default value is the id of the Kinect 2 class in the csra registry.
• --dr/--disable-registry. If this parameter is used, no input sources from the BaseCubeOne registry are used. This can be helpful for offline testing using transformation files.
• --sm-file-transformers is a list of scopes and transformation files which can be used for offline testing. Please refer to the code for the applicable file format. An input on the given scope will be transformed by the transformation specified in the file and then contributed to the merging process.

1.22.4.2. Ray Extractor¶

The Ray Extractor component extracts pointing ray distributions from the tracked postures merged in the Skeleton Merging component. These are sets of rays that belong to a single arm of a tracked posture. The sets can contain a single ray with the overall estimated pointing probability or multiple rays whose summed probability is the overall estimated pointing probability. To filter the ray distributions by a minimal overall probability, you can specify a parameter --re-threshold.

The ray extraction process occurs in two steps. First the probability of a pointing gesture is calculated for which several options exist which can be specified using the parameter --re-extractor:

• SIMPLE defaults to a probability of 1 times the certainty of the tracking of relevant joints and thus is unuseful for most applications.
• ARM_POSTURE works based on joint angles for a single arm that are classified based on a study-generated corpus (see CEORAUP for more details). Simplyfied it can be said that the more the hand is on chest level and the more the arm is straight the higher the calculated probability becomes. This strategy can be used to achieve a fast response by the system, as the probability is calculated based on a single frame. However, this configuration can lead to false positives in case of unintended short pointing gestures.
• POSTURE_DURATION Also uses the ARM_POSTURE strategy to calculate the probability of a pointing gesture for a single arm. However the probability is reduced by a factor and increases over time to its original value, if this is above a certain threshold. It is probably the most useful strategy as it can be configured in a way that suits every need from a fast response to a high avoidance of false positives. The configuration can be done using the following parameters:

• --re-duration-reduction-factor defines the factor by which the base probability is reduced initially. This factor will then be increased linearly to 1.0 over the lookback period.
• --re-duration-probability-threshold is the threshold which has to be exceeded by the base probability to cause an increase of the reduction-factor. If the base probability falls below the threshold, the reduction-factor will be reset to its initial value.
• --re-duration-max-angle defines the maximal angle difference in the shoulder-hand vector that may occur during the lookback period. If the angle is exceeded this indicates that no pointing gesture exists because there is too much movement.
• --re-duration-lookback is the period of time in milliseconds during over which the reduction factor is increased from its initial value to 1.0. If it reaches 1.0 it will remain at that value until the base probability drops below the probability-threshold or the arm is moved more than the max-angle inside a single lookback-period.

Note

Setting the --re-duration-reduction-factor to 1.0 makes the output equivalent to the ARM_POSTURE strategy

Multiple solutions have been implemented for this

1.22.4.3. Identification¶

The task of the Identification component is to determine what object pointing rays are directed at and with which probability a pointing gesture at a certain object occured. It receives collections of PointingRayDistributions and calculates UnitProbabilities based on these. Beside the default parameters, the following options are available:

• --identification-distance-measure takes one of the values ANGLE and ORTHOGONAL and selects the corresponding distance measure. Angular distance is the default which has worked well so far.
• --identification-unit-selector takes one of the values MAX and MEAN and defines in what way the probability results of a pointing ray distribution are reduced to a single value (Mean value over the rays vs. max of the rays). In combination with the other components and settings, only MEAN works at the moment, which is the default value.
• --identification-threshold is a probability threshold which can reduce the traffic for certain applications. If set to its default value 0.0, the probabilities for all relevant objects are sent via rsb on receiving pointing rays. Otherwise, the results are filtered using the threshold first.

1.22.4.4. Control¶

The Control component switches the power state of a unit, if it receives a UnitProbability with probability above a certain threshold on the selected-unit-scope. The probability threshold can be set using the parameter --control-threshold. Moreover, you can specify the duration in milliseconds which has to pass after a successful switch to be triggered again for the same unit using the parameter --control-cooldown-time. However, the default values of the program should work best in most scenarios.

1.22.5.1. Computer preparation¶

First of all you need to prepare a computer to be able to stream the Kinect data by following these steps (Don’t install the Kinect in the final location yet!):

1. Install Windows 8 or 10 from an image or freshly.
2. Kinect for Windows

1. Install the Kinect Studio and SDK for Windows (https://www.microsoft.com/en-us/download/details.aspx?id=44561)
2. Make sure that USB 3.0 is enabled (check the BIOS)
3. Connect a Kinect
4. Search the applications for SDK Browser 2.0 (Kinect for Windows) and run the Kinect v2 Configuration Verifier
5. Open Kinect Studio, go to File -> Settings and check Automatically connect on startup
6. Close Kinect Studio

3. Network access

1. Set a computer name
2. Register the computer’s mac adress to receive the same IP adress every time in the local net
3. Install VNCViewer
4. Connect the computer to the local net (192.168.75.121 pug.locnet & 192.168.75.120 apw.locnet)
5. Check whether you can access the computer using VNCViewer

4. RSB publishing

1. Copy the folder Tracking of the repository https://github.com/csra/kinect-pose-rsb-setup to C:\ and make sure, the folder C:\Tracking\Binary\ is registered in the PATH environment variable
2. Copy the folder C:\Tracking\config\ to the user directory of the current user (This contains the rsb config file)
3. Create a link to the file C:\Tracking\start_tracking.cmd on the Desktop and in the startup directory of the computer
4. Make sure bco is running and accessable via the DSC-computer
5. Run the file and wait a few seconds
6. Two windows should now be open: Kinect Studio and a console window with the following content:

Defined 1 streams!
1508861192884 rsb.depthsensors2.kinect2.kinect2device [WARN]: The Kinect2SDK only allows usage of one default sensor!
Waiting for termination signals is not implemented for this platform. The program will still terminate when such a signal is received but without shutting down properly.

If this is not the case:

• check network and rsb connection as well as the rsb.conf
• check the start_tracking.cmd file

7. From another computer run an rsb logger on scope /pointing/skeleton/COMPUTER_NAME where COMPUTER_NAME is replaced by the name of the computer and verify that data is published
8. Test that the last two points also happen automatically after a restart of the computer.

1.22.5.2. Placement¶

After the computer is set up and you tested connecting to it using VNCViewer and confirmed the correct output of the posture stream, you need to create a new Kinect 2 device config in the BaseCubeOne registry.

1. Note down the serial number of the Kinect
2. Install computer and Kinect in the correct location (Also check the camera view using VNCViewer and Kinect Studio)
3. Now place a marker on the ground in the camera’s field of view as described in Camera Calibration and take a screenshot using the program Color Basic - D2D of the SDK Browser 2.0 (Kinect for Windows).
4. Mirror the image along the x axis (Kinect takes mirrored images)
5. Follow the instructions in Camera Calibration to extract the camera location from the image also using the files: cameraPose.yml and markerSetConfig.yml from the repository https://github.com/csra/kinect-pose-rsb-setup.
6. Copy the location to a file which should have the following type of content (otherwise convert it):

RVEC: [-0.95784086, -0.11953047, 0.26125339;
  -0.2638337, 0.725905, -0.63518012;
  -0.11372176, -0.67732888, -0.72683764]
TVEC: [0.3816631;
  7.7323795;
  2.9691384]

Note

Don’t forget to add the location of the marker to the TVEC if necessary!

7. Run bco-psc-create-kinect -n NAME -p FILE -s SERIAL -l LOCATION where NAME is the name of the computer set in 3.1), FILE is the file created in the previous step,``SERIAL`` is the Kinect’s serial number on any computer connected to the live systems spread ring and LOCATION is the name of the location, that the Kinect should be placed in.
8. Check the results using the Registry Editor.
9. To improve the location (but not rotation) of the Kinect in the BaseCubeOne context, use the BComfy component (very carefully, to be as precise as possible).