Fused skeleton is not accurate in SDK 4.1

Software C++
1

Hello,
I’ve tested the sample “body tracking → multi-camera → cpp” with two Zed2 connected via usb.

I have seen that when a joint is not seen by both cameras the fused body is wrong, even if the raw skeleton of the camera that is seeing the body is right.
Here’s some example of what I mean.

zed3.PNG
zed3.PNG1626×1039 186 KB

zed4.PNG
zed4.PNG1255×1039 218 KB

zed1
zed11676×1040 302 KB

zed2.PNG
zed2.PNG1595×1039 222 KB

Green and Red skeletons are the raw skeletons detected from each camera and the Blue one is the fused skeleton.
You can see that the blue skeleton seems an average of the two raw bodies even if one of them is not showed because of its low confidence.

Can I set some parameter to get the joints of the fused body only from the showed raw skeleton joints (with high confidence) insted of considering every joint even if it has low confidence?

I tried to set different values of parameters from BodyTrackingFusionRuntimeParameters without success.

2

Hi @Rudiz,

Thank you for reaching out and reporting this issue.

Have you tried increasing the publisher detection_confidence_threshold to higher values?
Also, would it be possible to send two SVOs where you are able to reproduce the issue?
You can send them to support@stereolabs, in order for us to investigate the issue.

3

Hi @mattrouss,
sorry for my late reply.

I tried to increase and decrease the value of detection_confidence_threshold without success.
With higher values the problem is enhanced because the raw skeletons disappear more often.

I’ve just sent the SVO files via email.

I look forward to your response.
Thanks