Localization and Mapping for Autonomous Driving: Fault Detection and Reliability Analysis

Abstract

Autonomous driving has advanced rapidly during the past decades and has expanded its application for multiple fields, both indoor and outdoor. One of the significant issues associated with a highly automated vehicle (HAV) is how to increase the safety level. A key requirement to ensure the safety of automated driving is the ability of reliable localization and navigation, with which intelligent vehicle/robot systems could successfully make reliable decisions for the driving path or react to the sudden events occurring within the path. A map with rich environment information is essential to support autonomous driving system to meet these high requirements. Therefore, multi-sensor-based localization and mapping methods are studied in this Thesis. Although some studies have been conducted in this area, a full quality control scheme to guarantee the reliability and to detect outliers in localization and mapping systems is still lacking. The quality of the integration system has not been sufficiently evaluated. In this research, an extended Kalman filter and smoother based quality control (EKF/KS QC) scheme is investigated and has been successfully applied for different localization and mapping scenarios. An EKF/KS QC toolbox is developed in MATLAB, which can be easily embedded and applied into different localization and mapping scenarios. The major contributions of this research are: a) The equivalence between least squares and smoothing is discussed, and an extended Kalman filter-smoother quality control method is developed according to this equivalence, which can not only be used to deal with system model outlier with detection, and identification, can also be used to analyse, control and improve the system quality. Relevant mathematical models of this quality control method have been developed to deal with issues such as singular measurement covariance matrices, and numerical instability of smoothing. b) Quality control analysis is conducted for different positioning system, including Global Navigation Satellite System (GNSS) multi constellation integration for both Real Time Kinematic (RTK) and Post Processing Kinematic (PPK), and the integration of GNSS and Inertial Navigation System (INS). The results indicate PPK method can provide more reliable positioning results than RTK. With the proposed quality control method, the influence of the detected outlier can be mitigated by directly correcting the input measurement with the estimated outlier value, or by adapting the final estimation results with the estimated outlier’s influence value. c) Mathematical modelling and quality control aspects for online simultaneous localization and mapping (SLAM) are examined. A smoother based offline SLAM method is investigated with quality control. Both outdoor and indoor datasets have been tested with these SLAM methods. Geometry analysis for the SLAM system has been done according to the quality control results. The system reliability analysis is essential for the SLAM designer as it can be conducted at the early stage without real-world measurement. d) A least squares based localization method is proposed that treats the High-Definition (HD) map as a sensor source. This map-based sensor information is integrated with other perception sensors, which significantly improves localization efficiency and accuracy. Geometry analysis is undertaken with the quality measures to analyse the influence of the geometry upon the estimation solution and the system quality, which can be hints for future design of the localization system. e) A GNSS/INS aided LiDAR mapping and localization procedure is developed. A high-density map is generated offline, then, LiDAR-based localization can be undertaken online with this pre-generated map. Quality control is conducted for this system. The results demonstrate that the LiDAR based localization within map can effectively improve the accuracy and reliability compared to the GNSS/INS only system, especially during the period that GNSS signal is lost.