Map sensor fusion for lane keeping vehicle navigation
No Thumbnail Available
University of New Brunswick
The aim of this dissertation is to develop a lane keeping vehicle navigation system to bridge current challenges preventing safe, reliable, and scalable autonomous vehicle (AV) navigation. Current AV approaches are not globally scalable due to restrictions of the area of operation and cost. The reason is due to the use of machine learning techniques that can only operate where the training data was collected, and the use of expensive sensors, such as Light Detection and Ranging (LIDAR). Moreover, the lane keeping module of these AVs are mostly relying on imagery technology prone to failures in extremely bright or dark environments. With the lack of scalable and reliable lane keeping approaches, what set of sensors and approaches would provide an easy to scale, not limited in space, global lane keeping vehicle navigation? As this problem involves accurate tracking of the vehicle position on its lane of navigation in any environment, a reliable approach is a global navigation satellite systems (GNSS) inertial navigation system (INS) sensor fusion. These two sensors have a complimentary nature and can provide accurate solutions anywhere on the Earth’s surface. However, during long GNSS outages, INS alone cannot offer a reliable solution due to its inherent errors. We solve this limitation by providing a third piece of information, a map of the lanes. Maps have not been well explored in the vehicle navigation literature and, in this work, road maps it will play a crucial role in the navigation filter. Thus, with the intent of providing a global approach for a lane keeping vehicle navigation, a novel sensor fusion architecture combining a map of the lanes with GNSS/INS is proposed. This map of the lanes, obtained prior to navigation, precisely represents the location of the center of lanes which, with the development of a simple map matching algorithm, can be used to constrain the GNSS/INS filter. Two architectures were tested in two environments. The first experiment, a PPP with map constraints was tested in an open field. In the second experiment, a PPP/INS with INS/MAP integration was tested in an urban environment. From the first experiment, an improvement in PPP ionospheric-free residuals of 17.5 cm and 24.3 cm in average and standard deviation for pseudorange, as for the carrier-phase 3 cm in standard deviation with an unbiased average. By observing position displacement, PPP filter convergences and re-convergences were eliminated during GNSS outages. For the second experiment, position solution could be bridged with a loosely coupled INS/MAP filter, where INS drifts were estimated even during GNSS outages. An improvement in solution availability of 12% was obtained compared to a standard PPP filter. The cross-track error was within lane-level navigation based on proposed position accuracy limit. In both experiments the proposed architectures could track the vehicle position during long GNSS outages and increase solution availability and continuity. The main characteristics of this architecture are the independence of lighting conditions and easy to scale structure. The significance of these results is that current lane keeping systems can be strengthened by this architecture into potentially providing global autonomous navigation.