Hello, when using ZED for body tracking, when the person is partially occluded, the obtained data value is very unstable, and it manifests as swinging back and forth in the z-axis direction. And when this situation occurs, sl.BodyData.tracking_state is “OK” instead of other values. I have already tried modifying parameters such as depth_mode, prediction_timeout_s, detection_confidence_threshold, and minimum_keypoints_threshold, but I still can’t achieve a satisfactory effect.
May I ask what I should do to make the position data stable when there is occlusion?
Here is the demonstration video of the situation where data instability is caused by such occlusion.
Hello, I have tested the mode from ULTRA to NEURAL_PLUS. Only in the NEURAL_PLUS mode, the probability of position jitter is relatively low, and the jitter is likely to occur only once and then return to normal. However, it will still keep jittering in some cases of occlusion, and the occurrence of jitter will increase when the number of people increases.
As for other modes, it is basically very easy to find situations where jitter occurs.
In addition, after multiple tests, I adjusted the prediction_timeout_s to 0.5 and the detection_confidence_threshold to 50. It seems to be the best situation that can be achieved. However, I noticed that with these parameters, if one person passes by another person, the position will stay in place during the occlusion and will only move to the person after a short while. It seems that it is caused by the “seaching” state of sl.BodyData.tracking_state. And I’m not sure which parameter adjustment will affect the duration of this “seaching” state. I think it might be the “prediction_timeout_s”, but I can’t really feel the difference after making the modification.
Do you think there are any other parameters or methods that can make all of this even better? @BenjaminV
I’m wondering if adding a second sensor and using the Fusion function would help with the issue of people overlapping. I just ordered the Dual Camera Kit, and I hope I won’t have this issue with two sensors. @BenjaminV, what do you think?