Comparing the performance of different methods for estimation in inertial navigation systems

Navigation systems are systems that find the position, velocity and orientation information of an object with the help of a specific reference point. One of the most prominent of these systems is the Inertial Navigation System. Since this system works with its own internal system independent of external stimuli, its preference rate is increasing. Inertial Navigation System can estimate the position, orientation or velocity information of a vehicle with estimation algorithms by using data from accelerometer and rotometer sensors. In this study, the dataset consists of acceleration data of an automobile. These data were generated under variable speeds at a distance of 30.4 km on an intercity road. The measurements were obtained by running a MATLAB application on a smartphone. In previous studies, Kalman Filter algorithms have been predominantly used for estimation. In this thesis, in addition to the traditional Kalman Filter, the proven deep learning algorithms Long Short Term Memory (LSTM), Bidirectional Long Short Term Memory (BLSTM) and Gated Recurrent Units (GRU) are used as prediction algorithms. These deep learning algorithms were trained with different optimizers and the effects of the optimizers on the prediction results were also examined. Considering the findings obtained in the simulation studies detailed in the thesis, the most successful result belongs to the GRU deep learning network with an RMSE value of 2.5414. In LSTM and BLSTM deep learning networks, RMSE values of 2.5547 and 2.7592 were obtained respectively. In the studies conducted with the Kalman Filter, the RMSE was found to be 2.9322.