Enhancement of Fall Detection Algorithm Using Convolutional Autoencoder and Personalized Threshold

Falling is a hazardous situation for elderly people living alone and labor workers as they are easy to happen and can lead to serious injuries. Hence, a fall detection mechanism is an indispensable way to rescue victims without a delay. Various fall detection systems detect falling by using supervised deep-learning algorithms. However, labeling training data and collecting various falling motions data large enough for deep-learning is time-consuming and tiresome. Therefore, this study aims to develop a fall detection system utilizing the data not from falling but from usual motions of daily life. In this paper, an unsupervised learning method, convolutional autoencoder, and a wearable sensor, inertial measurement unit (IMU), were employed. The motion data from the IMU is converted to monochrome images for training and evaluating the developed fall detection algorithm. Falling is determined by comparing the input and output images of the model and a method for setting a threshold was investigated. After confirming the accuracy of the proposed method using a publicly available dataset and our dataset, the proposed method to train the model and to determine the threshold were addressed. Finally, the fall detection result with a sensitivity and a specificity of 100% and 99% was obtained.