Ego-localization Robust for Illumination Condition Changes based on Far-infrared Camera and Millimeter-wave Radar Fusion

Haruya Kyutoku, Tokihiko Akita, Seiichi Mita

Abstract: Vehicle ego-localization using in-vehicle sensors is one of the most important technologies for ADAS and AD. Accordingly, various attempts for accurate localization using in-vehicle sensors have been developed. Methods using a visible-light camera or a LiDAR have high accuracy, but they do not work properly in difficult environments such as rain, fog, and snow. In particular, visible-light cameras have a significant loss of information at night. On the other hand, farinfrared cameras and millimeter-wave radars are robust against difficult environments such as these. Therefore, we propose a robust ego-localization method using a fusion of a far-infrared camera and a millimeter-wave radar. The proposed method makes correspondences between in-vehicle far-infrared camera images and database images with correct location information using our epipolar geometry based metrics. Furthermore, the proposed method uses millimeter-wave radar information to solve a problem that this metrics have noises due to less texture information of far-infrared camera images. The experimental results show that the proposed method can estimate the position of the vehicle with about 30 cm error for the traveling direction under both day and night.