Learning Uncertainties in Ego-Motion Sensors for the Localisation of Autonomous Vehicles in GNSS-Deprived Environments

Abstract

In the absence of signals from Global Navigation Satellite Systems (GNSS), Inertial Navigation Systems (INS) are usually used to position vehicles. However, INS sensors are commonly plagued by noises, which grow exponentially over time during the multiple integration computation, leading to a poor navigation result. More so, the error drift is characterised by a time dependent pattern. This thesis proposes several efficient deep learning - based solutions to learn the position and orientation error drifts of the vehicle using Recurrent Neural Networks, such as the simple Recurrent Neural Network (sRNN), Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU) and the Input Delay Neural Network (IDNN). The thesis also investigates the use of the wheel encoder as an alternative to the accelerometer of the INS for vehicular positioning, and for the first time explores the potential of deep learning using the simple recurrent neural network to learn the uncertainty present in the wheel encoder’s output. These uncertainties could be manifested as wheel slips because of wet/muddy road drive or worn out tyres; or changes in the tyre size or pressure. The proposed solution has less integration steps in its computation, and as such, has the potential to provide a more accurate estimation of the vehicles position. In contrast to previous papers published in literature, which focused on travel routes that do not consider complex driving scenarios, this thesis investigates the performance of the proposed deep learning-based models to accurately estimate the position of the vehicle in challenging scenarios. These scenarios include, roundabouts, hard brake, sharp cornering, quick changes in vehicular acceleration, successive left and right turns. The performance of the deep learning models are then further evaluated extensively on longer-term GNSS outages of 30s, 60s, 120s and 180s duration respectively, over a total distance of 493 km. The experimental results obtained show that the proposed deep learning model using wheel odometry data is able to accurately position the vehicle with up to 93% reduction in the positioning error of its original (physics model) counterpart after any 180s of travel.

Date of Award	Dec 2021
Original language	English
Awarding Institution	Coventry University
Supervisor	Stratis Kanarachos (Supervisor), Vasile Palade (Supervisor) & Stavros Christopoulos (Supervisor)