Deep Residual Neural Network for Aggressive Physical Activity Recognition Using Surface Electromyography Sensors

Wearable sensors and machine learning have opened new doors for innovative automated systems. Smart wearables like smartwatches and wristbands can efficiently record human movements due to their integrated, compact sensors. This data collection is vital for understanding human behavior, a field known as human activity recognition (HAR), dedicated to classifying actions. Deep learning methods, which autonomously extract intricate features, have shown promise in HAR. Sensor-based HAR is extensively studied in academia and applied across domains like wellness tracking, medical diagnosis, and mobility analysis. However, challenges persist in detecting complex activities. Our study focuses on recognizing aggressive behaviors through surface electromyography (sEMG) sensor data analysis. We introduce ResNeXt, a novel deep neural network, to improve recognition accuracy. Through a series of tests, we evaluate ResNeXt's recognition capabilities and compare its results to other deep learning models. Our experiments reveal that ResNeXt outperforms other models, achieving an impressive 94.06% overall accuracy and an exceptional 97.68% accuracy in distinguishing normal from aggressive actions.