Direct Visual Localisation and Calibration for Road Vehicles in Changing City Environments

This paper presents a large-scale evaluation of a visual localisation method in a challenging city environment. Our system makes use of a map built by combining data from LIDAR and cameras mounted on a survey vehicle to build a dense appearance prior of the environment. We then localise by minimising the normalised information distance (NID) between a live camera image and an image generated from our prior. The use of NID produces a localiser that is robust to significant changes in scene appearance. Furthermore, NID can be used to compare images across different modalities, allowing us to use the same system to determine the extrinsic calibration between LIDAR and camera on the survey vehicle. We evaluate our system with a large-scale experiment consisting of over 450,000 camera frames collected over 110km of driving over a period of six months, and demonstrate reliable localisation even in the presence of illumination change, snow and seasonal effects.