LOCALIZATION OF PEOPLE IN GNSS-DENIED ENVIRONMENTS USING NEURAL-INERTIAL PREDICTION AND KALMAN FILTER CORRECTION

This thesis presents a method based on neural networks and Kalman filters for estimating the position of a person carrying a mobile device (i.e., cell phone or tablet) that can communicate with static UWB sensors or is carried in an environment with known landmark positions. This device is used to collect and share inertial measurement unit (IMU) information — which includes data from sensors such as accelerometers, gyroscopes, and magnetometers — and UWB and landmark information. The collected data, in combination with other necessary initial condition information, is input into a pre-trained deep neural network (DNN) which predicts the movement of the person. The prediction result is then periodically — based on outside measurement availability — updated to produce a more accurate result. The update process utilizes a Kalman Filter approach that relies on empirical and statistical models for DNN prediction and sensor noise. Therefore, the approach combines the principles of artificial intelligence and filtering techniques to produce a complete system which converts raw data to trajectory results of people. The initial tests were completed indoors where known landmark locations were compared with predicted positions. In a second set of experiments, GNSS location signals were combined with position estimation for correction. The final result shows the correction of neural network prediction with data from UWB sensors having known locations. Prediction and correction trajectories are shown and compared with the ground truth for applicable environments. The results show that the proposed system is accurate and reliable for predicting the trajectory of a person and can be used in future applications that require the localization of people in scenarios where GNSS is degraded or unavailable, such as indoors, in forests, or underground