Robust Self-Localization in the Urban Driving Environment

Abstract

Abstract A critical component to safe autonomous driving is an accurate state estimate of the host vehicle. This estimation relies on fusing data from onboard sensors, where a selflocalization algorithm utilizes available sensory information to produce an informed estimate of the vehicle’s state. However, the accuracy and availability of GPS signals, which heavily influences the self-localization algorithm, can be compromised in the urban driving environment due to various factors. This thesis addresses the self-localization problem that arises when sensor accuracy is non-stationary and affected by external factors, with particular focus on GPS-related faults. An Unscented Kalman Filter with asynchronous updates is employed, enabling the fusion of sensors operating at different sampling rates. The measurement model considers inputs from GPS, IMU, rotary encoders, and a pose estimate obtained from a visual SLAM algorithm. During filtering, the measurements are subjected to a normalized innovation squared test in order to detect and isolate faulty measurements. Furthermore, the system adapts to non-stationary sensor accuracy by a residual-based update of the expected measurement noise. The developed self-localization algorithm is evaluated using real-world sensor data collected by a research vehicle, where simulated GPS-related errors are injected in post-processing. The results demonstrate that the algorithm provides a robust state estimate, even in the presence of faulty GPS signals. The developed state estimator is proven to provide a more accurate positional estimate than the current state estimator of the research vehicle, which utilizes an Extended Kalman Filter, and also outperforms a conventional Unscented Kalman Filter. Additionally, subsets of the developed algorithm are evaluated to showcase the collective contribution of each component in achieving overall resilience to faulty sensor measurements. It is observed that the algorithm effectively mitigates the impact of three common error types: dropout, varying noise, and offset. However, it is also noted that not all types of drifting measurement errors can be effectively mitigated.