Counting stunt reps
A Little Problem
I do not know how many stunts I usually do in a stunt session
So I thought, lets find out! Should be easy enough.
The Goal
Count how many stunt sequences I do in a given session
TL;DR
On 2024-06-12 we did a 1,5h session. Just one base and one flyer.
Most of the session we did a variety of stunt sequences. Including a combination of tossing, walk ups, spinning, libs, QPs. Most of the sequences included multiple stunts, and a rare few, where only one skill was performed. To finish off we did Front Handspring Up Drills, focusing on the first part until the throw.
The stats:
- Grand total: 69
- Stunt sequences: 38
- FHSU drills: 31
Towards the solution
First steps
To achieve this, I could have just looked at all the videos and counted the repetitions this way. But if I can do it from video, then potentially, this can be automated. So I decided to take a different route. The first step ss to create some numeric signal.
To get a signal, I used a pose estimation model available in pytorch vision models.
Looking at the outputs
With the files processed I had a look. Analyzed a few signals to see if it is clear where there is a stunt in the file. The nose keypoint looked stable enough, so I plotted the Y and X coordinates for the nose keypoint.
The Y graph showed a clear signal for the stunt sequences, where there is clear movement up and down.
For the FHSU drill the X coordinate was actually better. As here the flyer goes to one side and then to the other. First the two lines are separate and then they come together on X graph, as the flyer approaches the base again.
Counting the sequences
I created plots for all the video files and then I just counted the number of stunts manually from the plots.
For this I created a CSV file with the filename and count as columns. And after counting all reps it looks like this.
| video_name | sequence_count | note |
|---|---|---|
| IMG_4776 | 5 | stunt sequence |
| IMG_4777 | 2 | stunt sequence |
| IMG_4778 | 3 | stunt sequence |
| IMG_4779 | 2 | stunt sequence |
| IMG_4780 | 1 | stunt sequence |
| IMG_4781 | 3 | stunt sequence |
| IMG_4782 | 3 | stunt sequence |
| IMG_4783 | 1 | stunt sequence |
| IMG_4784 | 4 | stunt sequence |
| IMG_4785 | 1 | stunt sequence |
| IMG_4786 | 4 | stunt sequence |
| IMG_4787 | 2 | stunt sequence |
| IMG_4788 | 1 | stunt sequence |
| IMG_4789 | 4 | stunt sequence |
| IMG_4790 | 2 | stunt sequence |
| IMG_4791 | 7 | FHSU drill |
| IMG_4792 | 4 | FHSU drill |
| IMG_4793 | 6 | FHSU drill |
| IMG_4794 | 6 | FHSU drill |
| IMG_4795 | 5 | FHSU drill |
| IMG_4796 | 3 | FHSU drill |
With the data available, we can just sum up the results.
And here we have it!
- Grand total: 69
- Stunt sequences: 38
- FHSU drills: 31
Other Observations
Along the way I made some extra observations.
Extra points on graphs
In some files there were these clear lines going across the graph.
So I looked at the image overlayed with keypoints.
Turns out tracked the people in the background walking by.
As the plotted keypoints were for all persons with score over 0.5, which is a fairly low value. We sometimes got traces of people in the background, clearly visible as the walked through the video frame. Even if the people were quite small.
Failed tosses
As we usually did longer sequences, then a failed toss becomes clearly different on the Y plot. Marked with red in the below image.
Different start distance in FHSU drill
In the FHSU Drill X-graph we can see that there are some reps, where the flyer was standing closer than usual.
On the image I marked two distances in green, and the red marking an approximate difference between the two.
Thats a wrap
Hope you found something useful.
If you have any idea how many stunts you do in a session, let me know. As of now I have no reference.