Localization for Autonomous Vehicles

Licentiate thesis, 2017

There has been a huge interest in self-driving cars lately, which is understandable, given the improvements it is predicted to bring in terms of safety and comfort in transportation. One enabling technology for self-driving cars, is accurate and reliable localization. Without it, one would not be able to use map information for path planning, and instead be left to solely rely on sensor input, to figure out what the road ahead looks like. This thesis is focused on the problem of cost effective localization of self-driving cars, which fulfill accuracy and reliability requirements for safe operation. In an initial study, a car equipped with the sensors of an advanced driver-assistance system is analyzed with respect to its localization performance. It is found that although performance is acceptable in good conditions, it needs improvements to reach the level required for autonomous vehicles. The global navigational satellite system (GNSS) receiver, and the automotive camera system are found to not provide as good information as expected. This presents the opportunity to improve the solution, with only marginally increased cost, by utilizing the existing sensors better. A first improvement is regarding global navigational satellite systems (GNSS) receivers. A novel solution using time relative GNSS observations, is proposed. The proposed solution is tested on data from the DriveMe project in Göteborg, and found capable of providing highly accurate time-relative positioning without use of expensive dual frequency receivers, base stations, or complex solutions that require long convergence time. Error introduced over 30 seconds of driving is found to be less than 1 dm on average. A second improvement is regarding how to use more information from the vehicle mounted cameras, without needing extremely large maps that would be required if using traditional image feature descriptors. This should be realized while maintaining localization performance over an extended period of time, despite the challenge of large visual changes over the year. A novel localization solution based on semantic descriptors is proposed, and is shown to be superior to a solution using traditional image features in terms of size of map, at a certain accuracy level.