A Wearable Fall Detection System Based on 1D CNN

Abstract

The current wearable fall detection systems mostly use threshold method with long-distance communication such as 3G/4G or machine learning algorithm with short-distance communication such as Bluetooth and Wi-Fi. But the former method has the problem of low algorithm accuracy, and the latter has the problem of short transmission distance. In order to solve these problems, an Arduino Nano 33 BLE development board with built-in accelerometer sensor is introduced. A deep learning model trained by 1D CNN (one-dimensional convolutional neural network) is trained offline firstly and transformed into a suitable model for the above development board using TensorFlow Lite. After deployment of a fall detection algorithm in an embedded terminal, the model has improved the fall detection accuracy. The inertial data is collected and normalized firstly and used as input data set for 1D CNN. The fall detection result and GPS data will be uploaded to the cloud using the NB-IoT (Narrow Band Internet of Things), and a warning message will be sent to the relative person. The fall accuracy of the above training model reached 98.85%, and the sensitivity and specificity were 98.86% and 99.84%, respectively.