A novel lightweight deep learning fall detection system based on global-local attention and channel feature augmentation

Abstract

Background and Objective: Reducing the number of falls in nursing facilities is crucial to prevent significant injury, increased costs, and emotional harm. However, current fall detection systems face a trade-off between accuracy and inference speed. This work aimed to develop a novel lightweight fall detection system that can achieve high accuracy and speed while reducing computational cost and model size. Method: We used convolutional neural networks (CNNs) and the Channel-wise dropout and global local attention module (GLAM) to train a lightweight fall detection model on over 10,000 human fall images from various scenarios. We also applied a channel-based feature augmentation module to enhance the robustness and stability of the model. Results: The proposed model achieved a detection precision of 95.1%, a recall of 93.3%, and a mean average precision (mAP) of 91.8%. It also had a significantly smaller size of 1.09M model parameters and a lower computational cost of 0.12 GFLOPs than existing methods. It could handle up to 20 cameras, simultaneously with a speed higher than 30 fps. Conclusion: The proposed lightweight model demonstrated excellent performance and practicality for fall detection in real-world settings, which could reduce the working pressure on medical staff and improve nursing efficiency.