RFAR: A Real-Time Firefighter Activity Recognition System Using Wearable Accelerometer

Human Activity Recognition (HAR) presents significant challenges. Although HAR is mainly used for elderly care, it holds significant potential for supporting firefighter operations. The main challenge of HAR is the recognition of complex actions and continuous transitions from real-time data. This study proposes a Real-time Firefighter Activity Recognition (RFAR) system that accurately recognizes eleven activities. The proposed system is based on a wearable device with an integrated accelerometer, ensuring compactness and energy efficiency. It can optimize parameters for deploying the random forest (RF) model on performance-constrained microcontrollers. A new data processing pipeline that integrates a sliding window technique with an abnormality index detection technique is proposed. This method achieves an F1-score of 96.1 % on a private dataset. Real-time experiments with a waist-mounted accelerometer achieved an F1-score of 96.2 % in rescue simulations conducted by volunteers at a firefighting training facility in Vietnam. To validate, the method is performed on three public datasets: achieving an F1-score of 96.6 % on SFDLA, 96.6 % on MobiFall, and 78.2 % on UniMiB-SHAR. Another contribution of this study is the development of a specialized dataset that comprehensively represents fundamental and unique firefighter activities in realistic scenarios.