Advancing human activity recognition with quaternion-based recurrent neural networks

Human activity recognition (HAR) stands as a vital nexus in the synthesis of healthcare, sports analytics, and human–computer interaction. This research introduces a groundbreaking approach to HAR by amalgamating the multidimensional strengths of quaternion algebra with the temporal sensitivity of recurrent neural networks, birthing the “Human Activity Recognition Utilizing Quaternion-Based Recurrent Neural Networks (QRNNs)” model. This innovative fusion targets the inherent challenges of high-dimensionality and temporal sequencing posed by wearable sensor data. The proposed QRNN model showcased promising results, achieving an accuracy rate of 98.46% after 20 training epochs, marking a significant advancement in HAR's state-of-the-art. The experimental results showcase the effectiveness and improved accuracy of HAR models with the utilization of quaternion algebra. Overall, this study offers an innovatiove way for wearable technology and human−machine synergy by ensuring an advanced mathematical and statistical framework for perceptual human activity identification.