Content area
Localization is critical for autonomous vehicles (AVs) to reliably estimate various AV states in multiple coordinate systems, such as the global pose in absolute UTM global coordinates, the ego velocity in the body frame and the lateral distance to lane boundaries on a map frame. AV tasks like navigation, path-planning and control rely on these state estimates in different coordinate frames for driving purposes. Numerous localization methods have been studied extensively to serve AV tasks with different localization outputs regarding the reference frame, accuracy level, update rate and latency. However, any single localization method using different data sources has its limitations and will encounter failures in dynamic real-world conditions. For example, methods using HD LiDAR maps are expensive, dead reckoning works only for short durations of time without an external fix, and others like GNSS fail for a variety of reasons.
For an AV to operate reliably and safely under a wide variety of conditions, its localization capabilities must therefore be able to utilize multiple correlated and complementary localization methods to obtain the best of all possible worlds.
This thesis proposes CORAL, COllaborative Relative and Absolute Localization, as a framework to meet these objectives. Collaboration in CORAL, working differently from conventional sensor fusion, receives multiple available localization outputs, associates them with other (correlated) outputs, and projects them to serve the needs of different AV tasks. Any individual localization failures are detected within one or a small number of processing steps. CORAL is made robust by using group theory that (a) formalizes the collaboration process with abstract algebraic objects that handle various localization outputs and their relationships homogeneously, and (b) scales extremely efficiently when new localization methods are introduced.
We propose two new localization methods Extended VINS-Mono (EVM) and LaneMatch to follow the notion of collaboration and to localize an AV in local, global and map reference frames with low-cost sensors and limited road information. EVM associates two relative localization methods (VINS-Mono and KF-based INS) with two absolute localization methods (GNSS and map-based ORB-SLAM2) to obtain multiple consistent (projected) absolute localization results to overcome GNSS interruption and disruption. LaneMatch associates an absolute localization method with a lane-matching technique to estimate the offset between the misaligned global and map coordinates at run-time to realize AV map-based localization.
Our two new localization methods and other off-the-shelf localization methods are integrated into our CORAL framework to form a localization system that provides AV state estimates for autonomous navigation, path-planning and control. On-road experiments demonstrate that CORAL efficiently enables collaboration among multiple localization methods, while also detecting and tolerating individual localization source failures.
