A Real-time Fall Detection System Using ToF Depth Images

Abstract

Fall accident is a major health problem resulting in serious injuries even deaths, especially among the elderly. Moreover, the population is aging globally, indicating that more fall accidents will probably happen in the future. A robust fall detection system is essential for avoiding injuries and deaths in fall accidents. In this paper, we propose a vision-based real-time fall detection system using Time-of- Flight depth maps with pose classification, height calculation, and temporal information processing. Our method sticks to the fall definition. We first divide the process of fall into three categories, and each subject in each frame is classified into different categories based on the pose by efficient YOLOv4 model. Combining with the depth data and temporal information in a time window, the proposed system can extract subjects' height and downward velocity, which are the key factors in fall judgment. Lastly the detection is achieved by comparing them with certain criteria. According to experimental results, our detection system achieved an average accuracy 97.92 % with a recall of 97.5 % and a precision of 98.3 % in off-line test on public dataset, and all three metrics above 100% in real-time test.