A Machine Learning Approach for Identifying Soccer Moves Using an Accelerometer Sensor

Human activity recognition is one of the active research topics in the past decade. Many applications have been developed in many domains such as health and sports. In this paper, we recognized main soccer actions using an accelerometer sensor, which was embedded in a mobile phone. The data was collected from 16 players. We extracted time and frequency domain features, tested five feature selection methods and eight supervised machine learning algorithms. We showed that Fast Hadamard Transform had a similar performance to Fast Fourier Transform, but Fast Hadamard Transform was more computationally efficient. In addition, we demonstrated that the recognition task can be achieved using two accelerometer axes instead of three axes. Support Vector Machine was the best classifier with an accuracy of 87% when all features were used. Using Wrapper method, a slight improvement in accuracy was achieved by Random forest to reach 88%. Using an Ensemble Learning approach, we were successfully able to achieve an accuracy of 90%.