WS-Fall: Fall Detection Attention-Based Approach using WiFi and Inertial Data Modalities

According to the World Health Organization, 37 million falls occur every year which leads to sever injuries that require medical intervention. In addition, more than 600,000 people end up dying due to falls making them the second cause of accidental injury fatalities worldwide. Specifically, the elderly are more prone to fatal falls. Therefore, fall detection systems are needed to allow medical professionals to intervene in such cases. Various wearable and non-contact fall detection devices have been developed recently. Since non-contact Wi-Fi-based methods are widely available and non-intrusive, they are gaining popularity. In this paper, we propose a state-of-the-art fall detection system that utilizes Wi-Fi and Inertial data which is independent of the environment and suitable for large-scale users. Our proposed WS-Fall detection model employs Convolutional Neural Networks with Gated Recurrent Units and couples it with an Attention mechanism, which allows the model to make use of both the spatial and temporal characteristics in the input data. The proposed model comprises three convolutional blocks, two GRU layers, an attention layer, and three fully connected layers. Our studies using the UT-HAR, MobiFall, and SisFall benchmark datasets showed that our WS-Fall fall detection model outperforms existing models by achieving an accuracy, precision, recall, and f1-score of 99.48% for the UT-HAR dataset, 98.53 for MobiFall dataset and 99% for the SisFall dataset.