Reducing false fall alerts in fall detection using deep learning models

Fall-related injuries pose a significant threat to older adults, with a 50 percent likelihood of mortality within six months if immobilized for over an hour after a fall. Effective fall detection and intervention strategies are imperative, yet distinguishing genuine falls from false alarms presents a challenge. The Center for Eldercare and Rehabilitation Technology at the University of Missouri, Columbia, addressed this challenge by deploying a fall detection system at the TigerPlace senior living facility in 2014. Despite advancements, false alarms persisted. In 2021, a supplementary analysis system utilizing Inception V3 and LSTM networks was introduced to further reduce false alarms. A key component of this system was the use of depth sensors instead of RGB cameras. Depth sensors were chosen primarily for privacy reasons, as they do not capture detailed visual images of individuals, thus minimizing concerns related to personal privacy and ensuring compliance with ethical standards. Despite these improvements, challenges remain due to false alarms triggered by various factors. To address this, the present study employs YOLOv5 for human detection in frames of the depth videos. The study includes preprocessing steps and runs Inception V3 and LSTM networks to verify accuracy and thresholds. The YOLOv5s model is applied to pre-processed videos, as well as the dataset generated after the LSTM model run. The outcomes of the LSTM+YOLOv5s, LSTM results and the YOLOv5s results are compared to evaluate the effectiveness of the LSTM+YOLOv5s model in detecting falls, and the results of the combination reduces the false fall alarms by 11 percent .The training dataset underwent manual annotation, and the resultant model accuracy was manually evaluated to ensure robustness, addressing privacy concerns comprehensively. This strategic approach aims to reduce false alerts and enhance fall detection systems, crucial for the well-being of elderly individuals. Achieving a balance between swift response and minimizing disruptions is essential for comprehensive fall management in aging populations.