Classification of physical exercises using a triaxial accelerometer in a smartphone and an artificial neural network

The prevalence of smartphones and their adequate computer skills can be used for detecting everyday physical exercises. Acquired information on performed exercises can be used in the field of Health Informatics. For identification of particular physical activity a number of sensors and their repositioning during exercises are needed. This paper presents a way to classify the type of exercise using only triaxial built-in accelerometric sensor in the smartphone. The smartphone itself is free to move inside the subject pocket. The problem of using a number of sensors and their repositioning during exercise is solved by raw signal filtering and by defining a set of signal descriptors. Nine characteristic exercises have been analyzed for different programs and levels of exercise. To filter the raw accelerometer signal a low-pass 10-th order Butterworth filter is used. The filtered signals are described in terms of five descriptors which are used to train an artificial neural network (ANN). Classification of the type of exercise is performed using ANN with an error of 0.7%. Some exercises can be performed with only left or right leg. The classification accuracy of proposed approach is tested in a way that the smartphone was always in the subject's right pocket even when the exercise is performed using left leg only.