Abstract
Roadside units that combine a millimetre-wave radar with a camera are becoming part of the road infrastructure, and using the two sensors together requires their extrinsic calibration. Above live lanes a target-based calibration is costly and has to be repeated, so calibration from the observed traffic, without targets, is the practical alternative. Any such method must first decide which radar track and which camera track belong to the same vehicle. Some published methods score each candidate pair from a pose fitted to that pair, which can be weakly constrained by a nearly straight vehicle track. All vehicles move on one road plane, so one plane-to-image homography must explain every correct match, and this paper builds the correspondence on that constraint. Calibration-free similarity cues rank the candidate pairs, the best-ranked pairs induce homography hypotheses, and each hypothesis is scored, without reference data, by the total cost of a global assignment of all radar tracks to all camera tracks. On eight hand-labelled recordings from one rooftop unit, the method assigns 656 of 672 radar tracks to the correct vehicle and selects the most accurate hypothesis on seven of them.