Illumination Invariant Imaging : Applications in Robust Vision-based Localisation , Mapping and Classification for Autonomous Vehicles

In this paper we propose the use of an illumination invariant transform to improve many aspects of visual localisation, mapping and scene classification for autonomous road vehicles. The illumination invariant colour space stems from modelling the spectral properties of the camera and scene illumination in conjunction, and requires only a single parameter derived from the image sensor specifications. We present results using a 24-hour dataset collected using an autonomous road vehicle, demonstrating increased consistency of the illumination invariant images in comparison to raw RGB images during daylight hours. We then present three example applications of how illumination invariant imaging can improve performance in the context of vision-based autonomous vehicles: 6-DoF metric localisation using monocular cameras over a 24-hour period, life-long visual localisation and mapping using stereo, and urban scene classification in changing environments. Our ultimate goal is robust and reliable vision-based perception and navigation an attractive proposition for low-cost autonomy for road vehicles.