Improving Height Estimation for Stationary Targets With 3-D Automotive Radar: From Uncertainty Analysis to Temporal Filtering

This article addresses the challenge of elevation angle estimation in 3-D automotive radar, a critical limitation for achieving accurate and reliable 3-D scene understanding in autonomous driving. While the vertical Doppler beam sharpening (DBS) provides a foundation for height estimation, existing implementations often suffer from limitations due to measurement noise. We enhance DBS using a rigorous uncertainty analysis and a robust, temporal filtering approach. Our analysis reveals the significant impact of target-sensor geometry, particularly small elevation angles, on estimation errors. To mitigate these uncertainties, we develop a simple yet effective method combining an extended Kalman filter (EKF) for temporal filtering with robust data association to reject spurious detections. Real-world experiments on highway and ITRI campus datasets, spanning 34 and 1.9 km, respectively, using a standard 3-D radar and a prebuilt LiDAR map for ground truth, demonstrate a substantial improvement in height accuracy. Compared with unfiltered DBS, our method increases height accuracy within 1 m from 53.41% to 62.32% on the highway and from 47.74% to 57.56% on the ITRI campus.