Learning a Better Attitude: A Recurrent Neural Filter for Orientation Estimation
Examensarbete för masterexamen
Systems, control and mechatronics (MPSYS), MSc
In the current paradigm of sensor fusion orientation estimation from inertial measurements unit sensor data is done using techniques derived with Bayesian statistics. These derivations are based on assumptions about noise distributions and hand crafted equations describing the relevant system dynamics. Machine learning, and more specifically neural networks, may provide an alternate solution to the problem of orientation estimation where no assumptions or handcrafted relationships are present. This thesis aims to investigate whether a neural network-based filter can achieve a performance comparable to, or exceeding that of, the more conventional extended Kalman filter. Two network architectures based on long short-term memory layers are proposed, trained, evaluated and compared using data from the Oxford inertial odometry dataset. Of the two suggested model architectures the socalled recurrent neural filter is found to give a the better performance. The recurrent neural filter has a structure inspired by Bayesian filtering, with a prediction and an update step, allowing it to output a prediction in the event of missing data. Further, the evaluated models are trained to estimate orientation as well as a parameterized error covariance matrix. Our results show that the suggested recurrent neural filter outperforms the benchmark filter both in average root mean square error and in execution time. The result also indicates that the machine learning-based approach for sensor fusion problems may be an attractive alternative to hand crafted filters in the future.
sensor-fusion , state estimation , absolute orientation estimation , recurrent neural filter , recurrent neural network , RNN , LSTM , IMU , MARG