Paper: Deep Learning for Safer Strength Training
Towards Safer Strength Training with Deep Learning for Rep Failure Prediction
TL;DR: Alongside Caterina Mammola ↗, we asked if strength training could be made safer. We show that a deep learning model is capable of predicting with (reasonable) accuracy when a rep in a bicep curl set is likely to be a failure. This technique could be expanded to excercises where muscular failure can lead to harm to reduce injury risk during unsupervised workouts.
We created a novel dataset of >3200 bicep curl reps performed to failure in and trained a Hierarchical LSTM that beats a linear baseline by nearly 30% at identifying the final safe rep. The model analyses joint motion patterns like range of motion and velocity across reps, inspired by Remaining Useful Life prediction from engineering.
View paper ↗
To decorate the read, the following are a few figures and videos that did not make it into the paper.
Our approach treats rep failure as a boundary detection problem, and we reframe it through the lens of Remaining Useful Life (RUL) prediction, adapting pose-based LSTMs to identify subtle biomechanical cues that precede failure.
We then run a pose estimation model on the video to extract the elbow angle over time.
Smooth the graph of elbow angle over time to remove noise and identify points that define the start and end of a rep.
While tweaking the parameter did not give failproof results (especially for sets with long pauses between reps or irregular reps) a rule based approach worked well.
To get participants to perform the reps, we used super advanced marketing techniques.
(In the small print we then explain that every set of reps is a ticket to a £40 prize draw, and that the more sets you do, the more tickets you get!)
We gathered 254 unique sets from 66 participants, totalling 3272 reps. A sizeable dataset but not enough to compare to other published approaches.
The grid shows a small subset of different reps from different sets and the data we extracted from them.
What follows is a large amount of tweaking deep learning approaches to improve its performance. We also developed a linear baseline that uses the same features as the model but linear.
Keep smiling,
Tomas