Pre-Impact Firefighter Fall Detection Using Machine Learning on the Edge

Falling is one of the primary causes of firefighter casualties. Therefore, early fall detection, followed by reliable and timely notification, is vital to ensure the survival of firefighters. However, due to complex and diverse firefighting scenarios, current fall detection systems (FDSs) are inadequate, especially in terms of accuracy and real-time applications. Hence, in this study, a wearable pre-impact FDS (PIFDS) for firefighters is proposed. We also evaluated the feasibility of using machine learning (ML) and ensemble learning (EL) methods deployed on the edge node to simultaneously improve the real-time performance and accuracy. Moving thresholding method is introduced to overcome the class-imbalanced issue due to fewer samples acquired during the pre-impact (PI) phase. Based on 14 firefighters' data collection, the experimental results revealed that a Decision Tree classifier with optimized parameters (DT-ED4) outperformed other ML and EL methods in a tradeoff between processing time and accuracy. The results also showed that our system can detect falls before the impact with an average lead time of 447.9 ms, sensitivity of 95.10%, and specificity of 97.99%. The system's heterogeneous IoT network architecture facilitated remote monitoring and alerts that enabled the incident commander (IC) to execute the rescue mission immediately upon notification. This study enhances the practicability of PIFDS and extends the application range of PIFDS from healthcare to firefighting.