Falls are one of the major health concerns faced by those 65 years old or over due to reduced mobility or coordination. Naveen Kumar understood this and decided to create a wearable fall detection device using an innovative Transformer model instead of traditional Recurrent Neural Networks (RNN) models for faster fall detection rates. Transformer-based neural network design was initially suggested in Vaswani and al’s 2017 article entitled, “Attention Is All You Need”. In 2017, the Transformer model used self-attention techniques to handle inputs simultaneously, which allows faster processing of long sequences compared to traditional RNN-based models. Transformer-based models have become the go-to architecture for various tasks in Natural Language Processing (NLP), such as machine translation (MT), language model creation and even text generation. In this article’s case, a Transformer-based model is employed to identify any changes in sensors collected by wearable devices.
Kumar knew his project needed to be fast while using minimal power, so he chose the Arduino GIGA R1 WiFi with its dual-core STM32H74XI Arm CPU featuring WiFi/Bluetooth(r) connectivity, along with being capable of interfacing with various sensors. Once he had connected the ADXL345 three-axis accelerometer, Kumar quickly realized that collecting all his data manually would take too long and take up too much of his time. So instead of downloading SisFall’s dataset and running an Python script to transform it into Edge Impulse-compatible format, he used an API to divide each sample into four-second segments and used Keras block editing feature to construct smaller Transformer models.
Training yielded a 202KB model, capable of accurately identifying 96% of falls that took place. Implementation involved using Arduino feature in sketch to run inference and display results on LEDs – however future versions could utilize GIGA R1 WiFi connection to send alerts when an accident was discovered; more details can be found here in Kumar’s article.
The post This GIGA R1 WiFi-powered wearable detects falls using a Transformer model appeared first on Arduino Blog.