비전 AI 기반 얼굴 및 행동 특징 융합을 통한 운전자 안전 모니터링 = Vision AI-Based Driver Safety Monitoring through Facial and Behavioral Feature Fusion

다국어 초록 (Multilingual Abstract)

This study proposes a universal real-time state recognition system based on multi-feature fusion using the MediaPipe framework. Real-time monitoring technology for human concentration and arousal states plays a crucial role in various fields including traffic safety, industrial safety, online education, and healthcare. A significant portion of accidents are caused by human factors such as drowsiness and distraction. The proposed system extracts facial landmarks and body landmarks in real-time using MediaPipe Face Mesh and Pose models, calculating Eye Aspect Ratio, Mouth Aspect Ratio, Head Pose angles, shoulder angle, and arm position. The multi-feature fusion module normalizes extracted features using Min-Max normalization and integrates them through weighted fusion with state-specific optimized weights. Sliding window and hysteresis mechanisms are applied for temporal consistency. Performance evaluation results demonstrate significant improvements over single-feature methods in drowsiness detection, distraction detection, and complex scenarios. This study overcomes the limitations of single-feature approaches through multi-feature fusion, achieving both high recognition accuracy and real-time processing performance. The weight-based approach ensures explainability, implementing a transparent system applicable to various fields. The system was designed as a universal framework applicable to multiple application areas rather than being limited to a specific application.

목차 (Table of Contents)

• Ⅰ. 서론 1
• 1. 연구 배경 및 필요성 1
• 2. 연구 목적 4
• 3. 관련 연구 6
• Ⅱ. 이론적 배경 8
• 1. MediaPipe 프레임워크 8
• 2. 랜드마크 기반 상태 인식 10
• 3. 멀티 피쳐 융합 기법 14
• Ⅲ. 시스템 설계 및 구현 16
• 1. 시스템 전체 구조 16
• 2. 운전자 모니터링 응용 20
• 3. 제조 현장 응용 시스템 22
• Ⅳ. 실험 및 결과 분석 24
• 1. 실험 평가 지표 및 결과 25
• 2. 유지 시간에 따른 성능 비교 30
• V. 결론 39
• 참고문헌 41