Geotagged Visual Localisation System for Urban Automated Vehicles 2022-01-0098

Vehicle localisation is one of the fundamental building blocks of automated driving systems. Although, high-end satellite navigation systems can provide centimetre-level accuracy, they are limited to applications where there is sufficient satellite signal visibility. One example where signals from navigational satellites might deteriorate is urban canyons which are characterised by high rise, high density residential and commercial buildings. To overcome the limitations of satellite navigation systems, most state-of-the-art localisation solutions fuse information from multiple sensors such as GNSS, LiDAR, camera, accelerometer, and wheel encoder with the purpose of creating a full 3D map of the operating environment. Although this approach provides accurate and reliable results, it is bounded in terms of data efficiency and scalability. With these limitations in mind an alternative methodology is proposed. More specifically, as opposed to existing approaches, the proposed system eliminates the need for creating full 3D maps by activating a visual localisation system only in geographical areas where the accuracy of satellite navigation systems might deteriorate, particularly urban canyons. The existence of urban canyons is pre-determined depending on the visibility of sky which is calculated using a digital surface model (DSM) of the environment. As a result, 3D maps are created only in challenging GNSS denied areas which makes the overall localisation and map systems much lighter improving the data efficiency and scalability. The paper will be covering the technical details of the proposed solution and will be showing the efficacy of the approach with results obtained in real-world urban environments.