Appearance-based vehicle localization across seasons in a metric map

Abstract

Great strides have been made in recent years in developing the necessary technologies to make autonomous cars a reality. However, a number of challenges remain, a major one being that of accurate vehicle localization. This thesis presents a vision-only approach to the outdoor localization problem. The system provides for real-time, metric localization of a moving camera (on a vehicle) in a pre-built 3D map, which is inherently robust with respect to appearance changes. This is achieved by utilizing a novel spatio-temporal map (STM) representation which is built up from multiple drives worth of stereo camera data, as well as a localization algorithm which efficiently retrieves landmarks from the STM to perform appearance-based localization in real-time. The STM encodes the landmark visibility structure of the datasets which were captured to build the map, as well as landmark descriptors and observation times. This visibility structure and meta-data are then exploited for efficient localization. In addition, by splitting the STM up into a number of submaps, computational tractability is ensured during the map-building phase, as well as during localization. Experiments on real data validate that the presented method works better than conventional approaches which operate in a map built of a single dataset.