Highly-efficient fog-based deep learning AAL fall detection system

Abstract

Falls is one of most concerning accidents in aged population due to its high frequency and serious repercussion; thus, quick assistance is critical to avoid serious health consequences. There are several Ambient Assisted Living (AAL) solutions that rely on the technologies of the Internet of Things (IoT), Cloud Computing and Machine Learning (ML). Recently, Deep Learning (DL) have been included for its high potential to improve accuracy on fall detection. Also, the use of fog devices for the ML inference (detecting falls) spares cloud drawback of high network latency, non-appropriate for delay-sensitive applications such as fall detectors. Though, current fall detection systems lack DL inference on the fog, and there is no evidence of it in real environments, nor documentation regarding the complex challenge of the deployment. Since DL requires considerable resources and fog nodes are resource-limited, a very efficient deployment and resource usage is critical. We present an innovative highly-efficient intelligent system based on a fog-cloud computing architecture to timely detect falls using DL technics deployed on resource-constrained devices (fog nodes). We employ a wearable tri-axial accelerometer to collect patient monitoring data. In the fog, we propose a smart-IoT-Gateway architecture to support the remote deployment and management of DL models. We deploy two DL models (LSTM/GRU) employing virtualization to optimize resources and evaluate their performance and inference time. The results prove the effectiveness of our fall system, that provides a more timely and accurate response than traditional fall detector systems, higher efficiency, 98.75% accuracy, lower delay, and service improvement.