관절 각도 기반 인체 자세 유사도 평가 기법

초록

Pixel-based pose evaluation is sensitive to environmental changes and fails to reflect actual poses, while large models like HRNet offer high accuracy but unsuitable for real-time use in resource-limited setting. This study builds pose vectors using MediaPipe joints and cosine similarity. Experiments show that angle-based evaluation offer higher discriminative power than pixel-based methods and achieves the highest FPS, demonstrating suitability for low-resource setting. Thus, the proposed method supports real-time dynamic pose analysis in constrained environments.

키워드