드론 기반 쓰러진 사람 탐지 성능에 대한 가상환경 데이터와 신뢰 스택 알고리즘의 영향 분석 연구

초록

The rising number of patients collapsing due to recent summer heatwaves has emerged as a significant public health concern. Although CCTV-based patient detection systems are available, they have limitations, including blind spots and installation challenges. This study develops and assesses an AI-based drone system for the real-time detection of fallen individuals. The proposed system integrates drone cameras with object detection technology to identify collapsed person. To overcome data collection difficulties, we incorporated Unity-based virtual environment data into our dataset. Additionally, a confidence stacking algorithm was implemented to reduce false positive rates and enhance detection accuracy. We evaluated the system's effectiveness by measuring detection accuracy using 2,500 real-world data samples and 10,000 virtual environment data samples. The experimental results show that the proposed system achieved an F0.5-score of 0.7812. The incorporation of virtual environment data and the confidence stacking algorithm result in improvements of 24% and 5% in the F0.5-score, respectively. These findings indicate that virtual environment data and confidence stacking algorithms significantly enhance the performance of fallen person detection systems.

키워드