Theory work
Five pose models and multiple public datasets were compared before choosing the best direction.
AI Sport Assistant
ABOUT PROJECT
WHAT IS THE PROJECT?
This project was created as my bachelor's thesis and focuses on developing a real-time system that uses computer vision to evaluate how correctly a person performs a sit-up.
WHAT WAS MY ROLE?
I handled the research, model comparison, implementation, testing, and evaluation of the system.
THEORY
MODEL AND DATASET RESEARCH
I researched multiple pose estimation models and datasets to compare accuracy, visualization approaches, and speed.
This helped determine the most suitable method for real-time exercise feedback.
PRACTICAL PART 1
RULE-BASED APPROACH
The first version detected sit-up quality using keypoint-based rules for upper-body movement and hand position.
Testing revealed challenges with hidden keypoints, side angles, and crossed arms.
MAIN RESULTS
Rule-based system
The first version worked but struggled with occlusion, camera angle variation, and more complex poses.
Improved model
Training on custom datasets significantly improved the accuracy of upper-body and hand-position detection.