-
Notifications
You must be signed in to change notification settings - Fork 65
Using The Data
When OpenPTrack is detecting and tracking people, objects, or poses it produces JSON tracks
and distributes them via UDP to the port and IP address listed in the configuration file opt_utils/conf/json_udp.yaml
. In addition to unicast behavior, broadcast or multicast can be configured in json_udp.yaml. Note that this file must be configured for your network for UDP output to be received on your client machine(s). (Also note that using multicast is not recommended unless you are familiar with how to configure and troubleshoot it.)
Each track update is published in a single UDP message with three possible payloads. One for person tracking, one for object tracking, and one for pose annotations. The following example is of a person tracking JSON message:
{
"header": {
"seq": 71251,
"stamp": {
"sec": 1415305737,
"nsec": 110138944
},
"frame_id": "world"
},
"tracks": [{
"id": 387,
"x": -0.89131,
"y": 2.41851,
"height": 1.55837,
"age": 29.471,
"confidence": 0.0500193
}]
}
- Units for x, y, and height are meters, which is true across all three JSON message types.
- ID is the unique identification associated with the track.
- Age is how long that ID has been active for.
- Confidence represents the system's calculation of how reliable the track is.
This is an example of object tracking's data format:
{
"header": {
"seq": 71251,
"stamp": {
"sec": 1415305737,
"nsec": 110138944
},
"frame_id": "world"
},
"object_tracks": [{
"id": 387,
"object_name": "ball_red",
"x": -0.89131,
"y": 2.41851,
"height": 1.55837,
"age": 29.471,
"confidence": 0.0500193
}]
}
- ID is again the unique identifier for this specific track.
- Object name represents the name associated to the object being tracking by the user in the MOT GUI.
- Units for x, y, and height are meters, which is true across all three JSON message types.
- Confidence represents the system's calculation of how reliable the track is.
Lastly, this is an example of the pose annotation's data format:
{
"header":{
"seq":71251,
"stamp":{
"sec":1415305737,
"nsec":110138944
},
"frame_id":"world"
},
"persons":
[
{
"id": 3,
"height":1.50029,
"orientation":2.4570,
"age":29.471,
"predicted_pose_name":"RIGHT_ARM_UP",
"predicted_pose_id": 2,
"prediction_score":0.988,
"poses":
[
{
"pose_name": "RIGHT_ARM_UP",
"pose_id": 2,
"prediction_score": 0.988
},
{
"prediction": "ARMS_UP",
"class_id": 1,
"confidence": 0.7
}
],
"joints":
{
"model":"rtpose_MPI",
"HEAD":{
"x":0.112,
"y":0.112,
"z":0.112,
"confidence":0.8971
},
"NECK":{
},
"RSHOULDER":{
},
"RELBOW":{
},
"RWRIST":{
},
"LSHOULDER":{
},
"LELBOW":{
},
"LWRIST":{
},
"RHIP":{
},
"RKNEE":{
},
"RANKLE":{
},
"LHIP":{
},
"LKNEE":{
},
"LANKLE":{
},
"CHEST":{
}
}
}]
}
- ID is again the unique identifier for this specific track.
- Orientation is an estimation of where the person being tracked is facing in relation to OpenPTrack's 0,0 coordinate.
- predicted_pose_name and pose_name represent the name associated to the object being tracking by the user during pose training.
- Units for x, y, z, and height are meters, which is true across all three JSON message types.
- predicted_pose_id and pose_id is the index number of the pose. This number is calculated starting at 0, which is the first training pose. Each new trained pose will be ordinal.
- prediction_score represent the system's calculation of how confident it is that the person being tracked is making the prose represented in pose_name and pose_id
- Then each of the 'joint' will have an x,y, and z score which is used to determine the pose.
If you are trying out face recognition, see this page for more information on the data format.
To view JSON tracking data on the host that is running the tracking process (generally, the master):
roslaunch opt_utils udp_listener.launch
The published rate defaults to 30 hz, but can be customized to suit your application by editing the rate
parameter in tracking/conf/moving_average_filter.yaml. UDP will then be published at an interval of your choosing.
Below, we have included code samples that demonstrate how to use this in various environments. If you want more detail about track lifetime, see track format detail
below.
If you need a simulator for OpenPTrack data, here is a start in Python: docs/assets/optsimulate.py, as well as a simple data receiver: docs/assets/optreceiver.py. Feel free to improve and submit pull requests for such samples!
Please see docs/assets/udp_example.py for a basic example of using Python to receive and parse OpenPTrack JSON data.
A simpler receiver example is also in docs/assets/optreceiver.py.
Please see docs/assets/node-ptrack.js for a basic example of using Node.js. All parsing is done in the client.
Please see docs/assets/max for an example of using OpenPTrack data in Cycling 74's Max/MSP.
See the operators released in 2017 by Ian Shelanskey.
One can use a Script DAT
to receive data and populate a DAT, or similarly for CHOP.
For an example of this in a drag-and-drop solution, use the touch component in docs/assets/openPTrack.tox.
It will publish an X, Y that is the average of all tracks. This can be modified as desired.
Please see docs/assets/p5_opt_udp.pde for a basic example of using Processing to receive, parse, and plot OpenPTrack JSON track data.
Please see Ros 2 NDN publisher for a ROS plugin to publish tracking data over NDN, as well as the consumer for receiving the tracking data. (More Information about NDN.)
ID is the person ID
. It does not wrap, and it has a 32 bit INT, so its maximum value is 2147483647.
There is a different behavior when a track has just been created (NEW). At first, a track is defined as NEW for sec_remain_new
seconds:
sec_remain_new.
During this period, a NEW track can be validated (and become NORMAL) if it gets at least detections_to_validate
detections in sec_before_fake
seconds:
detections_to_validate and sec_before_fake.
If a track does not become NORMAL within sec_remain_new
seconds, it is deleted. Once a track is NORMAL, it can be removed only if no detections are associated with it for sec_before_old" seconds
.
In general, an ID (track) is removed, and then no detections are associated with it for sec_before_old
seconds.
The sec_before_old
parameter can be changed here:
sec_before_old.
UDP data are also subject to the tracking filter
node, which adds a further parameter on top, that is track_lifetime_with_no_detections
:
track_lifetime_with_no_detections.
With this parameter, you can choose how long you want to see not visible
tracks in the UDP data (not visible
means with no detection associated at the current frame). If a track is not visible, what is published is its predicted position.
This parameter does not influence tracking, but only UDP data. Tracking can still continue to estimate where a person should be for sec_before_old
seconds (hoping to find it again), but you may want to remove tracks that are not visible for more than track_lifetime_with_no_detections
from visualization and the UDP data.
- System Requirements
- Supported Hardware
- Initial Network Configuration
- Example Hardware List for UCLA Setup
- Making the Checkerboard
- Time Synchronization
- Pre-Tracking Configuration
- Camera Network Configuration
- Single Camera
- Setting Parameters
- Multi-Sensor Person Tracking
- HOG vs YOLO Detectors
- World Coordinate Settings
- Single Camera
- Pose Initialization
- Multi Sensor Pose Annotation
- Pose Best Practices
- Setting Parameters
- Single Camera
- Setting Parameters
- Multi Sensor Object Tracking
- YOLO Custom Training & Testing
- Yolo Trainer
- Single Camera
- Setting Parameters
- Multi Sensor Face Detection and Recognition
- Face Detection and Recognition Data Format
How to receive tracking data in: