A Hybrid Multi-Person Fall Detection Scheme Based on Optimized YOLO and ST-GCN

Human falls are a serious health issue for elderly and disabled people living alone. Studies have shown that if fallers could be helped immediately after a fall, it would greatly reduce their risk of death and the percentage of them requiring long-term treatment. As a real-time automatic fall detection solution, vision-based human fall detection technology has received extensive attention from researchers. In this paper, a hybrid model based on YOLO and ST-GCN is proposed for multi-person fall detection application scenarios. The solution uses the ST-GCN model based on a graph convolutional network to detect the fall action, and enhances the model with YOLO for accurate and fast recognition of multi-person targets. Meanwhile, our scheme accelerates the model through optimization methods to meet the model's demand for lightweight and real-time performance. Finally, we conducted performance tests on the designed prototype system and using both publicly available single-person datasets and our own multi-person dataset. The experimental results show that under better environmental conditions, our model possesses high detection accuracy compared to state-of-the-art schemes, while it significantly outperforms other models in terms of inference speed. Therefore, this hybrid model based on YOLO and ST-GCN, as a preliminary attempt, provides a new solution idea for multi-person fall detection for the elderly.