Self-Localization for Autonomous Driving Using Vector Maps and Multi-Modal Odometry

Abstract

One of the fundamental requirements in automated driving is having accurate vehicle localization. It is because different modules such as motion planning and control require accurate location and heading of the ego-vehicle to navigate within the drivable region safely. Global Navigation Satellite Systems (GNSS) can provide the geolocation of the vehicle in different outdoor environments. However, they suffer from poor observability and even signal loss in GNSS-denied environments such as city canyons. Map-based self-localization systems are the other tools to estimate the pose of the vehicle in known environments. The main purpose of this research is to design a real-time self-localization system for autonomous driving.

To provide short-term constraints over the self-localization system a multi-modal vehicle odometry algorithm is developed that fuses an Inertial Measurement Unit (IMU), a camera, a Lidar, and a GNSS through an Error-State Kalman Filter (ESKF). Additionally, a Machine-Learning (ML)-based odometry algorithm is developed to compensate for the self-localization unavailability through kernel-based regression models that fuse IMU, encoders, and a steering sensor along with recent historical measurement data. The simulation and experimental results demonstrate that the vehicle odometry can be estimated with good accuracy.

Based on the main objective of the thesis, a novel computationally efficient self-localization algorithm is developed that uses geospatial information from High-Definition (HD) maps along with observation of nearby landmarks. This approach uses situation- and uncertainty-aware attention mechanisms to select “suitable” landmarks at any drivable location within the known environment based on their observability and level of uncertainty. By using landmarks that are invariant to seasonal changes and knowing “where to look” proactively, robustness and computational efficiency are improved. The developed localization system is implemented and experimentally evaluated on WATonoBus, the University of Waterloo's autonomous shuttle. The experimental results confirm excellent computational efficiency and good accuracy.