Body tracking accurancy

At my company we are using Kinect body tracking for recognizing and assessing fitness exercises. As the Kinect is no longer produced we urgently need an alternative. We are currently testing Orbbec among available alternatives, but it seems that the detection accuracy is not appropriate for our needs. I prepared a comparison of one of the basic exercises we have implemented - Biceps Curls between a Kinect and Orbbec here is the link to the video: biceps_curl_comparison.mp4 - Google Drive

As clearly visible the hand joints are lost in the upper position. The same problem occurs also with Kinect turned off so it is not the problem of two sensors interfering each other.

Are there plans to improve main detection algorithms? If so then when we can get the improvements?

1 Like

Hi pdondziak,

Thanks for these information. As the Orbbec Body Tracking SDK is currently under development and in an early beta release stage, some certain functions may not work as expected. Our team is working very hard to make improvements based on current feedback collected.
We desperately need such suggestions to be able to improve the algorithms for our future releases.
Please contact us directly at to provide your suggestions and ask for a survey to better helping us understand your requirements.

I already sent this survey with almost the same information that I posted here

Thanks for your support.

There is some clutter in the background (you can see the Orbbec SDK is detecting two extra bodies in the background). I’m wondering if that’s one of the sources of error.

This situation happen at most body tracking SDK so orbbec SDK too.
But I think it happens frequently more than other in orbbec SDK.
It will be improved.

Hi all!
I’m testing the orbbec body tracking on unity. I need it just to turn pages of a 3D book. The problem is that sometime it loose the hands when they are in front of the body.

Do you have some suggestion to avoid that? For example I’m using the orbbec astra at 45° in front of the user at 2.5 meters of height. Is there a better place?

There are a number of known issues with the skeletal system and location of hands - we have done extensive testing and found the OrbBec SDK way to unreliable for any sort of commercial work.

You might want to have a look at the Nuitrack SDK which we have found can produce better results. BUT even it too has issues with stability of the hand joints once they hands are in front of the body.

Seems like all the SDK developers think that T poses and star jumps are all we are looking to track = when the most common body poses for interaction are with the HANDS in front of the body. Something that all the current crop of SDK seem to fail at.

Whoever gets that 100% working wins our custom fwiw.


Same is true for us.
Using the Orbbec for research purposes is nice but commercially stable products have different requirements.
We have a Kinect-Sports-Training-System that works reliable and it seems that it will still take quite some time until Orbbec is commercially ready.
They will have to hurry or Project Kinect for Azure may be ready and take away the share that Kinect already has.

Well yes - the time is running out - Ive always said - the first group that gets a new TimeOfFlight sensor to market wins.

What people seem to forget is that Kinect2.0 was more like $1500 piece of hardware that MS all but gave away at a huge loss to drive people to their gaming platform.

Though if its Microsoft and bound to Azure - it may be a long climb uphill before it gets commercial acceptance