Robust Radar Mounting Angle Estimation in Operational Driving Conditions

TL;DR

This work addresses robustly estimating automotive radar mounting angles under real driving conditions where radar measurements are sparse and contaminated by moving objects. It proposes an odometry-based pipeline that fuses neural-network–driven radar motion estimates with an IMU yaw-rate model for bias and scale-factor compensation, and jointly estimates the mounting angle $\theta$ and inverse scale $s' = 1/s$ using a Taylor-expanded weighted least-squares formulation. The lateral-velocity equality constraint, expressed as $|V_t^{radar}| \sin(\beta_t + \theta) = (\tilde{\omega}_t / s) x_s$ (with $\tilde{\omega}_t = \omega_t - \bar{b}$), is aggregated over $T$ frames to solve for $[\theta', s']^T$ via $\mathbf{Y}=\mathbf{U}\mathbf{X}$ and weighted LSQ. Validation on the RadarScenes dataset (~79 km) demonstrates state-of-the-art accuracy and convergence within tens of seconds, without requiring controlled environments or dedicated targets, highlighting practical viability for in-vehicle calibration during normal operation.

Abstract

The robust estimation of the mounting angle for millimeter-wave automotive radars installed on moving vehicles is investigated. We propose a novel signal processing pipeline that combines radar and inertial measurement unit (IMU) data to achieve accurate and reliable performance in realistic driving scenarios. Unlike previous studies, the method employs neural networks to process sparse and noisy radar measurements, reject detections from moving objects, and estimate radar motion. In addition, a measurement model is introduced to correct IMU bias and scale factor errors. Using vehicle kinematics, the radar mounting angle is then computed from the estimated radar motion and the vehicle's yaw rate. To benchmark performance, the proposed approach is comprehensively compared with two problem formulations and four estimation techniques reported in the literature. Validation is carried out on the challenging RadarScenes dataset, covering over 79 km of real-world driving. Results show that the proposed method achieves state-of-the-art accuracy and robustness, with reliable estimates obtained within approximately 25 seconds of driving. To the best of our knowledge, this is the first study to demonstrate that automotive radar mounting angles can be accurately estimated in complex, real traffic conditions, without requiring controlled environments, dedicated targets, or specially designed driving routes.

Figures (6)

• Figure 1: The negative impacts of radar mounting angle misalignment on vehicle localization. The vehicle localization algorithm uses extrinsic parameters to convert the estimated radar trajectory into vehicle trajectory. However, when the extrinsic parameters are incorrect, additional errors will be injected after the conversion.
• Figure 2: Overview of the proposed signal processing pipeline for the problem of radar mounting angle estimation. Radar point clouds and IMU yaw-rate measurements are used as inputs, and the radar mounting angle is estimated as output. Radar motion is first estimated from the point clouds, while IMU yaw-rate measurements are modeled to account for bias and scale factor effects. The mounting angle is then obtained by enforcing a lateral velocity equality constraint within a kinematic formulation and solving a weighted least-squares problem. Here, $\mathbf{V}_t^{\mathrm{radar}}$ denotes the radar motion vector at timestamp $t$, $\omega_t$ is the vehicle yaw rate (rotational velocity), $(x_s, y_s)$ are the radar coordinates with respect to the vehicle rear center, $\beta_t$ is the direction of radar motion in the radar coordinate frame, and $\theta$ is the unknown radar mounting angle.
• Figure 3: Proposed signal processing pipeline for radar motion estimation. The estimator takes radar point clouds as input and outputs motion estimates and point weights. The weights are subsequently used to compute motion variance and reject sparse radar frames. Instead of RANSAC, two neural network-based approaches are employed: DeepEgozhu2023deepego for single-frame input, and DeepEgo+zhu2025deepego+, which incorporates temporal layers, for multiple frames.
• Figure 4: Density distribution of ego-vehicle speed and acceleration. For visualization, only 1% of the data (about 4.3k radar frames) was randomly sampled from the 64 selected recordings. This figure shows that the dataset used for evaluation covers a wide range of vehicle motion states during normal driving operations.
• Figure 5: Radar mounting angle estimation across 64 test scenes for different approaches. Solid blue: estimated angle per scene. Dashed blue: mean of all estimates. Dashed red: ground-truth mounting angle of Radar 3 from the RadarScenes dataset radar_scenes_dataset.