A multi-attribute fusion acceleration feature selection algorithm for activity recognition on smart phones

Current approaches for smart-phone based activity recognition focus mainly on the features extracted from the raw inertial accelerometer data, however, the raw sensor data is often heavily affected by the smart phone´s varying positions and orientations. Therefore, a feature selection algorithm for acceleration information based on multi-attribute fusion is proposed to overcome the influences. The algorithm fuses the mutual information, the difference between classes, the volatility in classes and the computation cost of selecting an optimal feature subset. The features in the subset are sensitive to the user´s activities, are not influenced by the smart phone´s positions. Experiment results show that the features selected by the proposed algorithm are suitable for activity recognition and have higher classification accuracy than the features selected by the traditional feature selection methods.