Automatic Fall Detection using Deep Neural Networks with Aggregated Residual Transformation

Abstract

The advancement of inertial sensor technology and the growing prevalence of intelligent wearables (including smart-watches, smart bands, and other intelligent devices) have aided in applying science and technology to automated Fall Detection Systems (FDSs). Over the past decade, wearable-based FDSs has garnered considerable scientific attention. In this context, ma-chine learning (ML) techniques have demonstrated remarkable efficacy in differentiating between falls and typical motions or Activities of Daily Living (ADLs) using the data recorded by wearable inertial sensors. Unfortunately, in most research, the effectiveness of machine learning classifiers was restricted by feature extraction and selection processes that relied on human-crafted decisions. A deep neural network was established to en-hance the capacity of fall detection, which combines convolutional layers and aggregated residual transformation in this study. The TensorFlow framework trained the proposed model to identify and categorize 15 special events: three distinct falls and twelve ADLs. On a publicly available benchmark dataset (UMAFall), the proposed model was analyzed and compared to existing baseline models. The findings indicate that the proposed model is superior to other models, with an accuracy of 96.72%.