Our project aimed to create an exercise-counter wearable based on a 6-axis IMU that would strap on their wrist, like a smartwatch, to count reps of various exercises performed. The motivation behind this was helping individuals at the gym keep track of their progression over time and receive detailed information about their workout trends. Current smartwatches allow one to track weight-lifting workouts but only display burnt calories by measuring heart rate over the course of the workout; they don’t count actual reps and exercises performed.
We approached the design considerations of this project using the BLERP model, discussed on the next page. For the scope of our project, we focused on 3 main exercises: Squat, Lat-Raise, and Crunch. We collected data by using IMU for these exercises with 2 extra state, idle and noise, trained a light weight neural net for edge devices through Edge Impulse, and deployed it on custom hardware using the Arduino Nano 33 BLE Sense board. The Training and Validation accuracy of our system was over 90%, and it worked exceptionally well in real life. There were a few shortcomings with the latency of the system which will be discussed further in the paper.
#Dataset ur data was almost exclusively collected by us. The primary contributors to our dataset were the 3 team members, and some of our friends. The dataset included men and women, and people of various heights. For actually collecting our data, we collected it by strapping the Arduino Uno to our wrists. We used a watch strap to hold the Arduino in place.
The data captured was Accelerometer, Magnetometer, and Gyroscope data in the x,y,z axis (9 data streams). Our data was 2 second long windows. Our final dataset was 12 minutes of 2 second long data with roughly equal samples for each of the 5 classes. This equated to ~72 samples of each label (12 mins * 60 seconds / 2 seconds per sample / 5 classes).
For more information, please contact the authors:
- saraltayal@cmu.edu
- davidcoo@andrew.cmu.edu
- ty2@andrew.cmu.edu