Fusion of activity recognition and recurrent neural network for attitude estimation improvement

Abstract

The paper focuses on machine learning-based processing of inertial measurement unit signals for attitude estimation. Signals from the accelerometer, gyroscope, and magnetometer are used as input to the trained machine learning models, based on the recurrent neural network. Models provides four quaternion parameters predicted by a pre-trained neural network. The practical application of such a system showed that it is hard to get a universal model that is suitable for precise attitude estimation on different types of activity. In this paper, a two-step solution is proposed, constructed from an activity recognition stage and switchable models for the prediction of quaternion parameters followed by attitude estimation. An experimental investigation was performed on publicly available data taken from the Berlin Robust Orientation Estimation Assessment Dataset. The tests, carried out with labeled data, showed that the preparation of activity-related quaternion parameter prediction models can decrease the mean error in attitude estimation by 12.5 % together with a reduction in standard deviation by 3.2 %.