Vision-based Fall Detection System: Novel Methodology and Comprehensive Experiments

Abstract

Falls are a significant health concern for older adults and people with disabilities. It can result in serious injuries such as hip fractures, head trauma, and even death. Therefore, a fall detection system is essential for preventing and mitigating the abovementioned negative consequences. In this study, we propose a fall detection system with a low-cost camera that can run in real-time and obtain persuading accuracy. Concretely, our proposed system first detects and tracks entities using the Yolo-V4-tiny model and SORT algorithm, respectively. Subsequently, the MediaPipe framework is utilized to extract the skeletons of each entity for physical characteristics calculations. Last but not least, the provided skeletons per frame with their corresponding physical characteristics serve as inputs of the Transformer to release detection results. The experimental result with reliable accuracy suggests that our study has the potential for practical applications. The demo video can be found here https://www.youtube.com/watch?v=dyFUT_SBrfU.