Vision Based Localization under Dynamic Illumination

Abstract

Localization in dynamically illuminated environments is often difficult due to static objects casting dynamic shadows. Feature extraction algorithms may detect both the objects and their shadows, producing conflict in localization algorithms. This work examines a colour model that separates brightness from chromaticity and applies it to eliminate features caused by dynamic illumination. The colour model is applied in two novel ways. Firstly, the chromaticity distortion of a single feature is used to determine if the feature is the result of illumination alone i.e. a shadow. Secondly, the chromaticity distortion of features matched between images is examined to determine if the monochrome based algorithm has matched them correctly. These two applications are put through a variety of tests in simulated then real world environments to assess their effectiveness in dynamically illuminated scenarios. The results demonstrate a significant reduction in the number of feature mismatches between images with dynamic light sources. The evaluation of the techniques individually in a Simultaneous Localization and Mapping (SLAM) task show substantial improvements in accuracy, with the combination of the two techniques producing a localization result that is highly robust to the environmental lighting.