Author(s): SHIEH, S., LIU, C., HSUEH, Y., Institution: NATIONAL YUNLIN UNIVERSITY OF SCIENCE & TECHNOLOGY, Country: TAIWAN, Abstract-ID: 595

INTRODUCTION:
This study focuses on the real-time motion detection of the Seoi-nage technique performed by judo athletes. The core visual engine of the motion analysis system adopts the Google MediaPipe Pose framework, whose underlying algorithm is based on the BlazePose lightweight convolutional neural network (CNN) architecture. Motion models from two athletes at National Yunlin University of Science and Technology in Taiwan, both gold medalists at the National Intercollegiate Athletic Games, were used to simulate multiple Seoi-nage executions and generate corresponding images. An efficient Detector–Tracker strategy was implemented, in which the system activates a detector during the initial stage to identify the human region of interest (ROI).
METHODS:
Seoi-nage is one of the most frequently used techniques in judo competitions (Gardasevic & Stankovic, 2019). To improve the accuracy of key joint localization, this study employs an encoder–decoder network and adopts a hybrid prediction strategy combining heatmaps and coordinate regression. Key points (hands, shoulders, and hips) are refined through regression for precise coordinate correction. Even in complex scenarios involving rapid limb crossings and frequent occlusions, the system consistently outputs 33 skeletal key points. Additionally, MediaPipe’s real-time computational capability and hardware accessibility enable the system to operate using a standard single-camera device and a laptop computer, effectively lowering the barrier to biomechanical analysis and allowing scientific data to be applied in grassroots teaching and research settings.
RESULTS:
The results indicate clear differences between the two athletes. The maximum instantaneous hand velocity of the first athlete was 81.4 pixels per frame, while that of the second athlete reached 89 pixels per frame. These findings demonstrate that the proposed method can effectively capture critical velocity changes during the execution of Seoi-nage and shows strong potential for real-time motion analysis and technical evaluation.
CONCLUSION:
Furthermore, instantaneous velocity analyses of the shoulder and hip points were conducted, leading to the identification of practical movement characteristics associated with higher speed, greater explosiveness, and improved efficiency.