Research on human fall detection based on lightweight YOLO and deployment with PyTorch mobile

To address the challenges of poor real-time performance and limited computational resources in mobile-based human fall detection, this paper proposes a lightweight YOLO-based algorithm that integrates spatiotemporal pose features. The model enhances YOLOv5s by introducing Ghost modules and depthwise separable convolutions for efficiency, along with a spatiotemporal attention mechanism to capture dynamic motion patterns. To further optimize deployment, a fourstage compression strategy (structured pruning, hybrid quantization, operator fusion, and hardware adaptation) reduces the model size to 3.5MB while maintaining accuracy. The system is deployed on mobile devices using PyTorch Mobile, achieving real-time detection with 91.7% accuracy on a Huawei Nova 11 SE. Experimental results demonstrate that the proposed method balances speed and precision, offering a practical solution for elderly fall monitoring. The study provides insights into lightweight model design and efficient edge deployment for real-world applications.