Multiscale Entropy for Physical Activity Recognition

This paper presents the evaluation of multiscale entropy (MSE) as a feature in physical activity recognition compared to the mostly used traditional features. Walking, jogging, and running were chosen as the physical activities for the comparison considering their similarities. Selection of similar activities can give a better evaluation of which features are useful in detecting slight differences. The acceleration data from x-, y-, and z-axes were collected using wearable accelerometers and then evaluated using Matlab and Weka. The k-Nearest neighbors (KNN), J48, and random forest (RF) were chosen as the classifiers. From the comparative evaluation, the MSE performed better compared to the traditional features. Further, the addition of the MSE significantly increased the performance of the traditional features.