비계 작업 근로자의 불안전행동 탐지를 위한 밝기 변화 스켈레톤 모델 개발

다국어 초록 (Multilingual Abstract)

The utilization of CCTV based on computer vision technology for safety management on construction sites enables supervisors to monitor the unsafe behaviors of workers from office settings and transmit information to on-site managers. However, the conventional use of CCTV for construction site safety management has limitations, requiring supervisors to be stationed in offices for continuous screen monitoring to identify workers' unsafe behaviors. To address these limitations, this study developed a skeleton-based video analysis technique capable of observing unsafe behaviors of construction workers at the level of localized body parts. The research focused on scaffolding installation sites and validates the performance of the developed technology through a comparison with existing artificial intelligence techniques. The validation involved testing the skeleton-based localized detection algorithm and detecting unsafe behaviors of workers in scaffolding installation sites under conditions such as worker presence, localized body exposure, and upper body exposure. Experimental results showed that the YOLO-based object recognition model achieved an accuracy of 86.4%, with 89.5% accuracy in detecting the trunk region of workers exposed outside the scaffold. In addition, it represented lower accuracy(53.8%) when detecting localized body parts. However, the skeleton-based object recognition model represented 94.9% accuracy, recognizing all instances of body exposure outside the scaffold and exhibiting a higher performance(76.9%) in detecting localized body parts. This study also analyzed the influence of brightness changes on worker and localized body part exposure detection, revealing that as brightness decreases, the accuracy of object recognition decreased. The YOLO model showed low accuracy in detecting localized body parts when obscured by objects like structures, while the skeleton model successfully detects exposed localized body parts and triggers exposure alerts. The skeleton-based object detecting technique was effective in detecting localized body parts, such as hands and arms, during work. Setting up danger detection zones along scaffold boundaries successfully triggers emergency alerted when a worker's body part was exposed outside the scaffold. Therefore, the utilization of skeleton-based object detecting technique in construction sites with scaffolding can reduce accidents and minimize fatalities by automatically observing and detecting unsafe behaviors of construction workers, even with limited safety management personnel.

목차 (Table of Contents)

• Ⅰ. 서 론 1
• 1. 연구의 배경 및 목적 1
• 2. 연구 방법 2
• 3. 연구동향 4
• 3.1 인공지능(Artificial Intelligence) 기술 4
• 3.2 스켈레톤 모델에 관한 연구 5
• Ⅱ. 이론적 배경 7
• 1. 기존 객체인식 모델 7
• 1.1 YOLO 7
• 2. 스켈레톤 모델 8
• 2.1 OpenPose 8
• 2.2 MediaPipe 10
• 2.3 YOLO-Pose 12
• 2.4 TensorFlow 13
• 3. 평가지표 15
• 3.1 정확도 (Accuracy) 15
• 3.2 Fps (Frame per second) 16
• 3.3 평균 정밀도 (mAP) 16
• Ⅲ. 스켈레톤 모델 개발 및 실험 20
• 1. 인공지능 기술 기반 불안전 행동 탐지 모델 개발 20
• 1.1 불안전 행동 탐지 절차 20
• 1.2 불안전 행동 탐지 알고리즘 선정 23
• 1.3 학습모델 생성 23
• 2. 실험계획 26
• 2.1 불안전 행동 탐지를 위한 데이터 세트 수집 26
• 2.2 영상처리 및 알고리즘 구현 26
• 2.3 밝기 변화에 따른 국소부위 객체인식 성능 비교 28
• Ⅳ. 불안전 행동 모델 평가 결과 30
• 1. 근로자 탐지 모델 결과 분석 30
• 2. 근로자 국소부위 탐지 모델 결과 분석 44
• 3. 밝기 변화에 따른 객체인식 성능 분석 56
• 3.1 YOLO 모델 56
• 3.2 MediaPipe 모델 71
• 4. 인공지능 모델 간 성능 비교 분석 86
• 4.1 불안전 행동에 따른 근로자 및 국소부위, 신체상부 노출 감지 86
• 4.2 밝기 변화에 따른 근로자 및 국소부위, 신체상부 노출 감지 87
• Ⅴ. 결 론 89
• 참고문헌 91
• 국문요약 94