FallNeXt: A Deep Residual Model based on Multi-Branch Aggregation for Sensor-based Fall Detection

Abstract

Falls are uncommon and pose a substantial health danger to adults and the elderly. These situations are a leading cause of severe injury. More harm could be averted if the faller could be located in time. The rising older population necessitates the rapid development of fall detection and prevention technologies. The burgeoning technology industry is focused on developing such technologies to improve the living conditions for the elderly in particular. A fall detection system monitors falls and provides an assistance notice to support mitigation of falls. This study proposes a sensor-based solution based on a deep learning network named FallNeXt to safeguard individual privacy and increase fall detection performance. This proposed network is a novel deep residual network that utilizes multi-branch aggregation to enhance fall detection capability. The detection effectiveness of this study was evaluated using three benchmark datasets for sensor-based fall detection: UpFall, SisFall, and UMAFall datasets. Compared to benchmark deep learning models on the three datasets, the experimental findings indicate that the proposed FallNeXt network scored the most significant overall accuracy and F1-score, with 96.16% and 99.12%, respectively. The benefit of the FallNeXt model's small but highly effective size for fall detection is its portability.