Fall Detection Based on Convolutional Neural Networks Using Smart Insole

In recent years, falls have become one of the leading cause of mortality for elderly people without caregivers at home. To tackle this problem, many autonomous monitoring systems for fall detection have been researched. However, considering the complicated computation during the manual feature extraction, these systems ignore the fact that the feedback to the guardian should be provided in real time. In this paper, we propose to apply one-dimensional (1D) convolutional neural networks (CNN) in automatic fall detection using built-in tri-axial accelerometer and gyroscope sensor in smart insole. By extracting signal features of the raw sensor data automatically, the proposed method aims to improve the fall detection rate with minimal computational complexity. Experimental results are provided to show the performance of the proposed system in fall detection and activity classification: on 800 falls and 1000 activities of daily living (ADL), the overall average accuracy of our method in detection can reach 98.61%. In addition, our system also achieves a high sensitivity and specificity of 97.92% and 99.58%, respectively.