A Study on Long Short-Term Memory Networks Applied to Local Positioning

Sammanfattning: Numerical approaches to lateration and sensor fusion are limited by inherent measure-ment errors and the positioning performance may benefit from alternative approaches. This thesis studies the applicability of deep learning to an Ultra Wide Band (UWB) based local positioning problem in a combination of readings from an Inertial Mea-surement Unit (IMU). Relying on the Robotic Operating System (ROS) and a robotic vacuum cleaner, sensor data was gathered and time stamped in the presence of a ground truth derived from a motion capture system. The gathered time series were processed and then used to train Long Short-Term Memory networks (LSTMs) for predicting two-dimensional coordinates and orientation in a plane. In a series of tests, accuracy and precision of the LSTM predictions were assessed and compared with two conventional approaches to positioning and orientation respectively. Results suggest that LSTMs can be applied well to positioning, however the study failed to establish benefits regarding orientation. It is concluded that the implemented LSTM increased positioning accuracy with 79.9 % and precision with 71.8 % compared to that of the conventional non-linear least mean squares approach. Comparing to the best recorded performance of the LSTM the on-chip sensor fusion with the utilised IMU was 15.8 %more accurate and 70.8 % more precise in estimating orientation from accelerometer, gyroscope and magnetometer readings. Despite this conclusion the study has found results indicating that significant improvements regarding orientating with LSTMs are within close reach.