Real-Time Radar--Vision Association via Monocular Distance Estimation

Benjamin Kiefer, Dominik Hildebrand, Rafia Rahim, Mahmut Karaaslan, Michael DeFilippo, Ersin Kaya, Andreas Zell

Keywords: maritime computer vision

TL;DR: Learning monocular distance stabilizes radar–vision association for maritime perception and enables reliable real-time fusion on embedded hardware.

Abstract: Associating camera detections with marine radar targets is challenging for autonomous surface vessels because monocular geometric projection provides reliable bearing cues but yields highly unstable distance estimates. In maritime scenes, small pitch errors, horizon ambiguity, and calibration drift make monocular range estimation an ill-posed problem, causing large distance errors that degrade downstream radar-vision association. We propose a real-time radar-vision matching framework that addresses this limitation by replacing geometric projection with a lightweight monocular distance estimator. The model predicts object range directly from image crops and replaces the geometric range component in the association cost, while retaining IMU-corrected projection for angular alignment. This produces more informative association costs and enables robust global matching via Hungarian assignment, even when sensor calibration or synchronization errors introduce bearing uncertainty. We implement the method on NVIDIA Jetson-class hardware and evaluate it on the MIT Sea Grant Marine Perception dataset with synchronized radar, RGB, and IMU data. The learned range estimator reduces distance error from 710\,m to 61\,m and consistently improves radar-vision association accuracy across multiple detector backends while maintaining real-time performance of 24 FPS on a Jetson Orin NX.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 20