Hi Hoon,
We filter the marker data at 12 Hz for activities detected as gait (walking and running), and just below 30Hz for all other activities. For activities like squatting, where the frequency content is below 30Hz, you could low pass filter at a lower frequency. In the paper, we filtered markers at 4Hz for squatting, based on Winter 1992. You could do this and re-run inverse kinematics. Alternatively, you could filter the kinematics directly at this frequency, which we do prior to dynamic simulations. See the preprint (
https://www.biorxiv.org/content/10.1101 ... 1.full.pdf) for details and rationale on how we filtered.
For more context to why there is some noise, there is no temporal consistency in the pose detection algorithms, so there tends to be frame-to-frame noise. The LSTM model that predicts anatomical landmarks tends to smooth the output marker positions. However, LSTM outputs are relative to the mid-hip keypoint, so the position of that keypoint is added back in to get global positions. Thus, there is more noise in the pelvis translations than other degrees of freedom. When we release the code, you will have the ability to re-run trials with more accurate pose detection settings in OpenPose, which tends to reduce this frame-frame noise. Filtering at frequencies that are common in the motion capture literature also works well for most low-frequency activities.
