Edge AI-Based Fall Detection for an Elderly Care Robot

Abstract

Falls represent a major safety risk for older adults, motivating reliable and privacy-preserving detection solutions that operate directly on edge devices. This work presents a lightweight fall-detection system based on the YOLOv5 object detector and deployed on an NVIDIA Jetson Nano to recognize three postures: standing, sitting, and falling. When a fall is confirmed, the embedded system generates an alert that can be transmitted to a caregiver. On a separate test set, the detector achieves a mean Average Precision of 95.8 percent at an Intersection-over-Union threshold of 0.5. Real-time measurements on the Jetson Nano show an effective processing speed of around seven frames per second for the baseline configuration and up to twelve frames per second when using NVIDIA’s TensorRT optimization. The software pipeline is configured for compatibility with PyTorch, CUDA, and TensorRT, and is adapted to the constraints of embedded deployment. Current limitations relate to lighting variations, camera viewpoint, and scene complexity. Planned developments will focus on improving robustness, reducing latency, and integrating the system on an elderly-care robot for on-board perception and timely alerts.