This paper explores the possibility of localizing an automotive-radar-equipped vehicle within an urban environment relative to an existing map of the environment created using data from visible light cameras. Such cross-modal localization would enable robust, low-cost absolute localization in poor weather conditions based only on radar even when the vehicle has never previously visited the area. This is because a pre-existing absolutely-referenced visible-light-based map (e.g., constructed from Google Street View images) could be exploited for localization provided that a correspondence between features in this map and the vehicle’s radar returns can be established. The greatest challenge presented by cross-modal localization with automotive radar is the extreme sparseness of automotive-radar-produced features, which prevents application of standard computer vision techniques for the cross-modal registration. To the best of the authors’ knowledge, cross-modal localization using automotive- grade radar within a visible-light-based map is unprecedented. The current paper demonstrates that it can be used for vehicle localization with horizontal errors below 61 cm (95%).

Cite and download the paper:
Peter A. Iannucci, Lakshay Narula, and Todd E. Humphreys, "Cross-Modal Localization: Using Automotive Radar for Absolute Geolocation within a Map Produced with Visible-Light Imagery," In 2020 IEEE/ION Position, Location and Navigation Symposium (PLANS), pp. 285-296. IEEE, 2020.