Anomaly detection for integrating lidar and satellite vehicle localization data

Sometimes, when localizing a vehicle, there are cases where one of the sources gives unreliable data due to unforeseen conditions that arise. For example, when driving into a garage, the global satellite navigation system may cease to generate reliable data; when leaving the existing 3-dimensional map of the area, reliable lidar localization ceases to be received; and at night, localization using on-board camera data is difficult. In all these cases, combining multiple localization data sources with anomaly detection can help. Localization defines the position of the vehicle and the angles of its orientation in three-dimensional space. This article presents the developed AnKF lidar and satellite navigation data integration method based on the extended Kalman filter and anomaly detection algorithms. Particular attention is paid to the study of classical and neural network approaches to detecting anomalies in multidimensional time series, which consist of detecting and processing deviations in data streams obtained during the localization of an unmanned vehicle. To quantify the results, a labeled data set was created based on the CARLA unmanned vehicle simulator. The paper shows that the detection of anomalies in the behavior of the system can significantly improve the quality of localization of an unmanned vehicle.