Developing a GA-optimized EWMA feature engineering method for real-time human activity recognition

Developing a GA-optimized EWMA feature engineering method for real-time human activity recognition

Real-time human activity recognition (HAR) is garnering attention across various fields, such as healthcare, fitness and sports, security and surveillance, occupational safety, smart environments, and more. This is largely attributed to the rapid development of mobile devices, which enable users to record human activity signals using accelerometers. In this study, we found that the recognition rates were poor when tri-axial activity signals collected from accelerometers were directly fed into classifiers, including decision trees (DT), discriminant analysis (DA), logistic regression (LR), Naïve Bayes classifiers, support vector machines (SVM), ensemble learning (EL), and neural networks (NN). The recognition rates improved from 75% to 94% when the three-axis signals were transformed into statistical signal features (SSF). Despite the improvement in accuracy, the increase in the number of input variables from 3 to 66 has burdened the computation time. Furthermore, a higher recognition rate is needed to have an effective decision making. Therefore, this study develops a novel feature engineering method by using genetic algorithm (GA) and exponentially weighted moving average (EWMA). The EWMA is not only used to capture the characteristics of time sequences derived from the activity signals but also to eliminate redundant SSFs. GA is employed to optimize EWMA weights for each SSF. The results demonstrate that the Ensemble Bagged Trees classifier, using the proposed GA-optimized EWMA features, achieves a testing recognition rate of 95.2% with a prediction time of less than 0.01 s, making it suitable for the field of real-time HAR.