Ultra low-power, wearable, accelerated shallow-learning fall detection for elderly at-risk persons

Area of Science:

• Wearable technology
• Biomedical engineering
• Machine learning for healthcare

Background:

• Unintentional falls pose a significant risk to the elderly due to diminished physical capabilities.
• Existing fall detection methods may lack accuracy or real-time prediction capabilities.
• Need for unobtrusive, low-power, and reliable fall detection systems for at-risk populations.

Purpose of the Study:

• To develop and manufacture a wireless, wearable, low-power fall detection sensor (FDS).
• To predict and detect falls in elderly individuals at risk.
• To enhance the safety and well-being of older adults through timely and accurate fall detection.

Main Methods:

• Utilized low-power field-programmable gate arrays (FPGAs) to implement a fixed-function neural network.
• Employed a Convolutional Neural Network (CNN) model trained using the Caffe deep learning framework.
• Integrated an ST Microelectronics LSM6DSOX inertial measurement unit (IMU) sensor with an ultra-low-power Lattice iCE40UP FPGA.
• Collected and published a dataset of 3D accelerometer and gyroscope measurements from human subjects performing activities of daily life (ADLs) and falls.

Main Results:

• Developed a functional wireless, wearable, low-power fall detection sensor.
• Demonstrated the capability of the CNN model on FPGA for activity categorization, including fall detection.
• Acquired and published a novel dataset for fall detection research.
• Successfully accelerated convolutional operations on a machine learning model for ultra-low power FPGA deployment.

Conclusions:

• The developed FDS offers a promising solution for enhancing elderly safety.
• The integration of CNNs on FPGAs enables efficient, low-power fall detection.
• The published dataset will facilitate further research and development in the field.
• This innovative approach contributes to proactive fall prevention strategies for at-risk populations.