Physical fitness is vital for overall well-being, impacting cardiovascular health, cognitive focus, obesity prevention, and mortality risk. This thesis explores real-time exercise recognition and tracking to enhance the workout experience, providing immediate feedback by recognizing and counting repetitions. This approach optimizes exercise effectiveness and reduces the risk of injury due to incorrect posture or technique, motivating individuals to push their limits and improve engagement with fitness routines.
While smartphones and fitness trackers enable activity tracking, they often struggle to recognize various exercises reliably and incur significant costs. Consequently, artificial intelligence (AI) has become crucial, penetrating diverse industries, including fitness. Human pose estimation, a key aspect of computer vision research, has gained prominence in this context.
This thesis introduces two pose estimation methods, YOLOv7 and MediaPipe, and conducts a comparative analysis between the two through illustrative examples. Following this, our attention turns to building a machine learning model for a virtual personal fitness system. This system possesses the capability to recognize exercises and count repetitions in real-time using a camera. Opting for the MediaPipe pose estimation model, we leverage the MediaPipe library for keypoint extraction and data collection, while TensorFlow powers the machine learning aspect. Specifically, we employ recurrent neural networks, particularly long short-term memory, within a Python software environment to develop the entire model. The culmination of this work presents compelling statistical data, showcasing an impressive 100% classification performance.
|
