Correct exercise execution and accurate repetition counting are essential
for safe home workouts, yet existing solutions rely on markers or computationally
heavy deep models that are not robust on mobile devices and
suffer from high latency. This thesis presents real-time, markerless bodypose
classification with a lightweight k-nearest-neighbours (k-NN) classifier
over normalized BlazePose embeddings, reliable repetition counting via EMA
smoothing and hysteresis, and an on-the-fly assessment of execution quality.
The approach is embedded in Fitness Devil, an Android application that,
at a chosen time, locks the phone until the user completes a camera-verified
workout. Development was guided by a 50-participant user study, and an
economic analysis confirms the project’s viability and its potential for business
(B2B) expansion.
|
