Comparing the Performance of Different Methods for Estimation in Inertial Navigation Systems

Abstract

There are many positioning systems available today. The most prominent of these systems is the Inertial Navigation System, which is increasingly preferred because it works with its own internal system independent of external stimuli. This system de-tects position, orientation and velocity information by means of accelerometer and gyroscope sensors. Using this information, it is possible to make predictions for the next position, orientation and speed with various algorithms. In the studies conducted so far, Kalman Filter (KF) algorithms have been predominantly used for prediction. In this study, Long Term-Short Memory (LSTM) neural network architecture, Bidirectional Long Short-Term Memory (BLSTM), Gated Recurrent Unit (GRU) and Kalman Filtering methods, which are among the deep learning algorithms that have proven themselves as prediction algorithms, are examined in detail and a comparative study is presented. Here, LSTM, BLSTM and GRU deep learning networks were first trained with IMU sensor data and speed estimation was performed. Root Mean Squared Error (RMSE) values were obtained as 2.5547, 2.7592 and 2.5414, respectively. Furthermore, the same deep learning network methods were trained with GPS data. The prediction data obtained through LSTM, BLSTM and GRU provided RMSE values of 0.42542, 1.91122 and 0.32274, respectively. We see that the prediction with GPS data have higher accuracy since deep learning networks trained with GPS were less affected by noise during the training phase.