Pose detection for human motion characterisation using AI, in a gym setting
Project Abstract
Exercise provides multiple health benefits, such as increased strength, muscle mass and bone mass. However, current exercise tracking methods, such as fitness apps or wearables rely on manual input, which can be time consuming and inaccurate. This is especially true for strength training where reps are often the unit of measurements when performing a workout. My dissertation proposes an exercise tracking system that utilizes Human Pose Estimation and object detection used in conjunction to detect reps, provide real-time form correction, while reducing injury risk. Specifically, MediaPipe Pose is used to extract body coordinates, while an object detection model identifies the equipment being used. To answer whether a machine learning system can accurately track resistance exercises, while offering form feedback, tests are performed on dumbbell curl, barbell bench press, and deadlifts, as these exercises form the foundation of many strength training programs. Each exercise is tested on video data, with 10 repetitions performed. The system performed with almost perfect accuracy when counting reps, however correct form feedback proved difficult when testing on a head on camera angle.
Keywords: exercise biomechanics, ai, machine learning
Conference Details
Session: Poster Session A at Poster Stand 133
Location: Sir Stanley Clarke Auditorium at Tuesday 7th 13:30 – 17:00
Markers: Daniele Cafolla, Julian Hough
Course: BSc Computer Science, 3rd Year
Future Plans: I’m undecided