For those aged 65 and over, falls can be one of the most serious health concerns they face either due to lower mobility or decreasing overall coordination. Recognizing this issue, Naveen Kumar set out to produce a wearable fall-detecting device that aims to increase the speed at which this occurs by utilizing a Transformer-based model rather than a more traditional recurrent neural network (RNN) model.
Because this project needed to be both fast and consume only small amounts of current, Kumar went with the new Arduino GIGA R1 WiFi due to its STM32H74XI dual-core Arm CPU, onboard WiFi/Bluetooth®, and ability to interface with a wide variety of sensors. After connecting an ADXL345 three-axis accelerometer, he realized that collecting many hours of samples by hand would be far too time consuming, so instead, he downloaded the SisFall dataset, ran a Python script to parse the sample data into an Edge Impulse-compatible format, and then uploaded the resulting JSON files into a new project. Once completed, he used the API to split each sample into four-second segments and then used the Keras block edit feature to build a reduced-sized Transformer model.
The result after training was a 202KB large model that could accurately determine if a fall occurred 96% of the time. Deployment was then as simple as using the Arduino library feature within a sketch to run an inference and display the result via an LED, though future iterations could leverage the GIGA R1 WiFi’s connectivity to send out alert notifications if an accident is detected. More information can be found here in Kumar’s write-up.
