Improving Accuracy On Lower Body Rapid Movements

[Ryan Shaffer]

Hello,

I am having difficulty capturing rapid martial arts movements like roundhouse kicks. The current specs my setup uses are 2-3 iPhone 14 recording at 240 fps using the OpenCap's full body model. I saw in a previous post to use HRNET with the shoulder model for upper body kinematics, I was wondering what your recommendations were for improving lower body kinematics though? I have posted GIFs of unsuccessful and successful trials below. These trials include a successful front kick and an unsuccessful roundhouse kick. The trials in this post were with a 2 camera setup. My team has experimented with 2 and 3-camera setups. We found that the 2 camera setup was much better than a 3 camera setup for capturing the front kick, but we haven't been able to capture a roundhouse kick properly like the front kick at all no matter the setup. We are unsure what to do as possible occlusions don't seem to be the entire reason why we can't get proper data for a roundhouse kick. We have yet to use OpenCap core to rebuild the failed trials. We are going to use the hand synchronization technique that was previously overlooked as well in the new trials. Thank you for your help, what are your recommendations?

I could not upload GIFs of the trials due to their size but if you copy the links below you can see them in another window

The front kick successful with 2 cameras, trial at 0.5 speed:
https://www.kapwing.com/exports/65678f6 ... ee0f1fc6ee

The round house kick unsuccessful with 2 cameras, trial at 0.5 speed:
https://www.kapwing.com/exports/656790e ... bcc90affa0

[Scott Uhlrich]

Can you verify that you were using the v0.3 augmenter model? It has been default for the past ~month. If you download your data, the augmenter model will be in the session metadata.

If so, then it can be helpful to visualize the marker data in the OpenSim gui to see if errors are coming when creating the triangulated keypoints or the augmented ones (*_study).

[Ryan Shaffer]

Hi Dr. Uhlrich,

We set our recording setting to v0.3 model, HRnet, the improved shoulder model, and 120fps. Additionally, we used the punching the hand in the air technique at the beginning of the trials to use the complex movement algorithm. This has greatly improved our results, thank you so much for your help!