My current research focus involves extracting keypoint coordinates from video frames of exercise movements, such as squats, barbell lifts, and overhead presses. I am utilizing the Temporal Shift Module (TSM) to detect speed changes in these movements. I am conducting experiments with various backbone models, including MobileNetV2, ResNet50, ResNet101, HRNetw32, and HRNetw48, to design a model that minimizes latency and maintains performance for real-time operation on mobile devices. A key aspect of my research is finding the optimal backbone model that balances PCKh and latency, while also investigating methods to enhance overall performance

Our backbone model MobileNet V2 demonstrates a PCKh of 86.278 and a latency of 34.5ms when tested on a Galaxy Note 8. The objective of this research is to improve the AI model used for analyzing exercise motions to achieve a PCKh of 90 or more and reduce latency to under 30ms. To ensure practical application, I will evaluate the model’s performance and usability in real-world scenarios, incorporating user feedback for continuous refinement. Additionally, I intend to expand the exercise motion dataset, developing a model robust to the inherent variability of real-world data. Finally, I am committed to the goal of releasing the developed AI fitness application in October 2025