Automatic classification of physical exercises from wearable sensors using small dataset from non-laboratory settings

Effective classification of physical exercises allows individuals to assess their levels of physical activity and functional ability for maintaining physical fitness and help reduce risks of chronic diseases. This paper investigates and compares classification techniques for detecting physical exercise in real-world contexts that often only supports a small training dataset. The system combines heart rate with other exercise-related features, such as distance, duration, calories, etc. The experiment uses a dataset of 40 realistic (uncontrolled) sessions from 22 individuals wearing wearable sensors while performing different exercises, including walking, aerobics, running, indoor cycling, and weight training. Based on a 5-fold cross validation approach, AdaBoost demonstrated the highest (87.25%) classification accuracy compared to other classifiers, including support vector machine, neural network, and binary decision tree when used individually. When fused together at the decision level using majority-voting techniques, these classifiers achieved higher accuracy (89.25%) than that of individual applications.