Lightweight Skeleton-Based Fall Detection for Real-Time Edge Vision Systems

Abstract

Falls among elderly individuals pose significant health risks and require timely and reliable detection. This paper presents a lightweight vision-based fall detection approach based on skeleton-derived posture features for efficient edge deployment. Human pose estimation is performed using a TensorFlow Lite MoveNet model to extract skeletal keypoints, from which three geometric features - torso angle, body aspect ratio, and floor proximity, are derived to characterize fall-related postures. These features are combined using a threshold-based logical decision rule, enabling an interpretable and training-free classification mechanism suitable for resource-constrained devices. The approach is designed to maintain low computational complexity while achieving robust detection performance. Experimental results show that the skeleton-based feature model achieves 98% accuracy, 95.2% precision, and 100% recall in distinguishing fall and non-fall activities. Additionally, an event-triggered activation mechanism is incorporated to support efficient edge operation by reducing unnecessary processing. The results demonstrate that the method provides an effective and practical solution for real-time fall detection in smart elderly care environments.