Lightweight fall detection on edge devices via knowledge distillation

Abstract

Falls pose a significant risk to the elderly, often leading to severe injuries and medical emergencies. Timely and accurate fall detection is crucial for effective intervention. While vision-based approaches offer high accuracy and non-intrusive monitoring, they typically require large-scale deep learning models, making deployment on resource-constrained edge devices impractical due to high computational demands. To address this challenge, we propose an efficient fall detection framework using knowledge distillation to transfer knowledge from a high-performance teacher model to a compact student model. This approach significantly reduces model complexity without sacrificing accuracy. We apply cross-background and cross-person validation for robust evaluation. Experimental results show our model improves F1-score by up to 7% while requiring only 1/200 of the teacher’s parameters, making it suitable for real-time edge deployment.