TinyFallNet: A Lightweight Pre-Impact Fall Detection Model.

Area of Science:

• Gerontology
• Computer Science
• Biomedical Engineering

Background:

• Falls are a major health risk for the elderly, necessitating effective detection methods.
• Existing deep learning fall detection models require optimization for deployment on microcomputer units (MCUs).
• Wearable airbag systems require efficient preimpact fall detection algorithms.

Purpose of the Study:

• To develop a lightweight deep learning model for preimpact fall detection using inertial measurement unit (IMU) data.
• To benchmark the state-of-the-art ConvLSTM model and explore lightweight alternatives.
• To validate the proposed model's performance on elderly fall data and activities of daily living.

Main Methods:

• Leveraged features from VGGNet and ResNet image-classification models to create a lightweight deep learning model.
• Developed and evaluated models using the KFall public dataset with IMU data from young subjects.
• Proposed TinyFallNet based on ResNet, optimizing for memory efficiency while maintaining accuracy.
• Validated the algorithm using elderly fall data from the FARSEEING dataset and KFall ADLs data.

Main Results:

• TinyFallNet achieved 97.37% accuracy, slightly lower than ConvLSTM's 98.00%, but required significantly less memory (0.70 MB vs. 1.58 MB).
• Demonstrated the successful application of image-classification models for IMU-based preimpact fall detection.
• Confirmed the potential for further lightweighting deep learning models through data type-specific tuning.

Conclusions:

• TinyFallNet offers a viable, memory-efficient solution for preimpact fall detection suitable for MCUs.
• The study highlights the adaptability of image-classification architectures for IMU data processing in fall detection.
• This research contributes to lightweight deep learning models for IMU-based applications, including wearable safety devices.