비디오 기반 포즈 추정을 이용한 실시간 축구 드리블 기술 분류를 위한 머신러닝과 딥러닝 모델의 비교 분석

초록

This paper proposes machine learning models for classifying five soccer dribbling techniques using video-based pose estimation. We collected 48 videos of dribbling techniques (feinting, inside-outside, flip flap, Ronaldo chop, step over), extracted 3D coordinates of 33 anatomical landmarks from 21,186 frames using MediaPipe Pose, and compared six models through 5-fold cross-validation. XGBoost achieved the highest accuracy (99.11%), followed by Random Forest (98.94%), while deep learning models showed lower performance: 1D CNN (97.64%), LSTM (97.31%), Transformer (88.03%), and GRU-CNN (86.66%). All models exceeded real-time requirements with XGBoost reaching 203,484 FPS. Feature importance analysis revealed lower body landmarks contributed 68% to classification decisions. Results demonstrate tree-based ensemble methods outperform deep learning for dribbling classification by effectively utilizing spatial coordinate relationships, enabling real-time, high-accuracy analysis for training applications.

키워드