Instrumentation & Measurement Magazine 24-6 - 25

Fig. 3. Synchronization effect on training.
less time to train on average. We end up with the following 2-L
configuration: Input-42, Hidden-88, Hidden-26, Output-28.
Post-Processing
This function merges both the coordinates from the pretrained
model and the coordinates predicted by the SNN using
the pre-trained model's confidence score using the Weighted
Coordinate Function (WCF). Another way to see this function
is in Fig. 4: the more confident the original prediction is, the
more similar the fixed coordinate is. If the confidence score is
low, it will be more similar to the SNN prediction.
Feedback Generation
The PoseNet model gives us the keypoints in terms of x and y
coordinates. However, since no two people have the exact proportions
and will never be in the same position and location
within the video frames, we need to perform post-processing
work to use this information. Both the coach and the trainee
videos go through the same post-processing before comparing
their poses.
There are four steps in this process:
◗ Confidence Level Filtering: Since we will not have a reasonable
estimation of all the keypoints, we need to use
only those with enough confidence level. We will use
keypoints that have 80% confidence or more.
◗ Origin Keypoint Calculation: We proceed similarly to [11]
by creating local and global parameters. We use three
origins instead of only one to help in occlusion cases,
where some keypoints could not be estimated because
they were covered. Our global parameters are the angles
between a keypoint and three origin points: nose, shoulders,
and hips.
◗ Angular Parametrization: The collection of angles for each
of the relevant joints (wrists, elbows, ankles, and knees)
to the parent (local parameters) and the origin keypoints
(global parameters) is called the Angular Pose Representation.
The advantage of using the APR to compare
both the coach and the trainee poses is two-fold: first, it
is independent of body dimensions, which means that
we can compare human bodies of different height, arm
length, etc.; and the second advantage is that the users
can be at any location in the video plane, as we are not
using coordinates anymore, but angles between the same
body points to measure the pose. In Fig. 5, we can see the
keypoint " right wrist " local and global parameters.
◗ Scoring: We obtain the score by calculating the difference
between the coach pose representation angles and the
trainee pose representation angles. The score is presented
to the trainee in real-time while they are performing the
routine. This allows the user to correct any pose that they
Table 1 - RSO results
Layer Count
1
2
3
4
5
10
September 2021
Top 10 MSE
Average
0.0177
0.0157
0.0140
0.0131
0.0129
0.0156
Top 1
MSE
0.0176
0.0150
0.0136
0.0126
0.0123
0.0149
MSE Improv
0%
17%
10%
8%
2%
-18%
Top 1 #
Params
7585
8660
21038
19399
25429
29614
IEEE Instrumentation & Measurement Magazine
Param
Incr
0%
14%
143%
-8%
31%
16%
Time per
sample
(us/sample)
183.58
179.86
195.40
214.56
189.70
213.34
25
Instrumentation & Measurement Magazine 24-6

Table of Contents for the Digital Edition of Instrumentation & Measurement Magazine 24-6
