Robust Graph SLAM in Challenging GNSS Environments Using Lidar Odometry

Abstract: Localization is a fundamental part of achieving fully autonomous vehicles. A localization system needs to constantly provide accurate information about the position of the vehicle and failure could lead to catastrophic consequences. Global Navigation Satellite Systems (GNSS) can supply accurate positional measurements but are susceptible to disturbances and outages in environments such as indoors, in tunnels, or nearby tall buildings. A common method called simultaneous localization and mapping (SLAM) creates a spatial map and simultaneously determines the position of a robot or vehicle. Utilizing different sensors for localization can increase the accuracy and robustness of such a system if used correctly. This thesis uses a graph-based version of SLAM called graph SLAM which stores previous measurements in a factor graph, making it possible to adjust the trajectory and map as new information is gained. The best position state estimation is gained by optimizing the graph representing the log-likelihood of the data. To treat GNSS outliers in a graph SLAM system, robust optimization techniques can be used, and this thesis investigates two techniques called realizing, reversing, recovering (RRR), and dynamic covariance scaling (DCS). High-end GNSS and Lidar sensors are used to gather a data set on a suburban public road. Information about the position and orientation of the vehicle are inferred from the data set using graph SLAM together with robust techniques in three different scenarios. The scenarios contain disturbances called multipathing, Gaussian disturbances, and outages. A parameter study examines the free parameters Φ in DCS and the p-value in the RRR method. The localization performance varies less when changing the free parameter in RRR than in DCS. The localization performance from RRR is consistent for most values of p. DCS shows greater variation in the localization performance for different values of Φ. In the tested cases, results conclude that Φ should be set to 2.5 for the most consistent localization across all states. RRR performed best with a p-value set to 0.85. A lower value led to too many discarded measurements which decreased performance. DCS outperforms RRR across the tested scenarios but further testing is needed to determine whether RRR is better suited for handling larger errors.