A sedentary lifestyle is a major risk factor for chronic diseases such as cardiovascular diseases, diabetes mellitus and osteoporosis. Many studies reported that physical activity can prevent such pathologies. Thus, measuring physical activity has become increasingly important to conduct research and develop tools aimed at disease prevention. Several researchers have approached the challenge of recognizing human activities such as walking, running and cycling from wearable sensors. A trade-off is often observed between complexity of the classification problem and user-friendliness of the measurement system. We investigated a template matching-based framework to recognize sport activities using one single accelerometer placed at the wrist. User-independent signal templates were created and five distance and correlation-based matching techniques were proposed for activity classification to target eight sports activities such as cycling, cross-trainer, rowing, squatting, stepping, running, walking and weight lifting. The framework was evaluated using data collected in a gym environment involving two different groups of volunteers: normal weight and overweight subjects. The viability of our classification approach was analyzed by comparing its recognition accuracy with that of four popular classifiers: Decision Tree, Naïve Bayes, Logistic Regression and Artificial Neural Network. Based on this comparison, we concluded that template matching is well suited for the recognition of sport activities due to their inherent periodic nature. In particular, our template matching-based method showed to be robust to data generated by previously unseen subjects with different biometric characteristics and possibly motor skills.
Keywords: Activity classification, Dynamic Time Warping,Template prototypes, Overweight subjects