Good Thinking

Computer vision system tracks workouts when wearables can't

View 2 Images
GymCam automatically detects and counts repetitive body movements
VitalikRadko/Depositphotos
GymCam can simultaneously track multiple users performing different exercises
Carnegie Mellon University
GymCam automatically detects and counts repetitive body movements
VitalikRadko/Depositphotos

Many people already use wearables such as smartwatches to count their repetitions while exercising. Researchers at Carnegie Mellon University have developed what could be a better alternative, however, in the form of a user-monitoring computer vision system.

The problem with wearable devices is that while they're able to identify exercise-associated movements of the body part upon which they're being worn, they may miss others. It was with this limitation in mind that PhD students Rushil Khurana and Karan Ahuja created the GymCam system.

It utilizes an ordinary stationary video camera, that's hooked up to a computer – the technology could even be incorporated into a smartphone app. As the camera "watches" the user while they're working out, a custom-designed algorithm automatically detects repetitive body movements, matches them up to known exercises, and counts the number of reps.

And although the system can be used by individuals in their homes, it's also capable of tracking many users at once within the same room. In fact, by learning which exercise machines are located in which parts of a room, the system is better able to determine which activity each person is performing.

GymCam can simultaneously track multiple users performing different exercises
Carnegie Mellon University

Obstacles such as the gym equipment itself, which may partially block the camera's view of the user's body, reportedly aren't problematic when it comes to tracking and identifying movements. And because the system doesn't record or transmit images of users' faces, their privacy is protected.

In tests that were performed in a crowded university gym, the current version of GymCam was reportedly 84.6-percent accurate at differentiating exercises from other sorts of movement, 93.6-percent accurate at then recognizing the type of exercise being performed, plus it was able to count reps within an accuracy of plus or minus 1.7.

Khurana and Ahuja are presenting their research this Thursday in London, at the International Joint Conference on Pervasive and Ubiquitous Computing.

Source: Carnegie Mellon University

  • Facebook
  • Twitter
  • Flipboard
  • LinkedIn
0 comments
There are no comments. Be the first!