Human Activity Recognition by Wearable Sensors : Comparison of different classifiers for real-time applications

In recent years, there is a growing interest in Human Activity Recognition (HAR) systems applied in healthcare. A HAR system is essentially made of a wearable device equipped with a set of sensors (like accelerometers, gyroscopes, magnetometers, heart-rate sensors, etc…) and a classifier able to recognize the activity performed. In this study we focused on the choice of the classifier, since there isn’t a unique and consolidated methodology for HAR. The main aim of the study is to compare the performances of 5 classifiers, based on machine learning. Furthermore, we analyzed advantages and disadvantages of their implementation onto a wearable and realtime HAR system. We acquired magnetic and inertial measurement unit (MIMU) signals from 15 young volunteers. For each subject, we recorded 9 signals from tri-axis accelerometer, gyroscope and magnetometer. All signals were divided in 5s-windows and processed to extract 342 features in time, frequency and time-frequency domains. By means of two feature selection steps (correlation-based and genetic algorithm), we reduced the number of features to 69. These features were used as input for the following 5 classifiers: K-Nearest Neighbor (KNN), Feedforward Neural Network (FNN), Support Vector Machines (SVM), Naïve Bayes (NB), and Decision Tree (DT). Our results showed that all classifiers were able to correctly recognize more than 90% of activities. The best performances were obtained by KNN. Analyzing advantages and disadvantages of each classifier for its implementation by means of a microcontroller the most suitable was DT. In fact, this classifier can be easily implemented, it has low memory and computational requirements, and it allows for a further reduction of the required features.