CREVE: An Acceleration-based Constraint Approach for Robust Radar Ego-Velocity Estimation

TL;DR

CREVE addresses robust ego-velocity estimation from sparse mmWave radar by introducing acceleration-based inequality constraints derived from IMU acceleration. The method adds an adaptive gamma rule, tied to the radar inlier ratio, and an accelerometer bias estimation routine to maintain acceleration fidelity over time. Empirical results on IRS DoerIros2021 and ColoRadar show CREVE significantly improves absolute trajectory error and maintains real-time performance against REVE, DeREVE, and RAVE. The approach enhances radar-inertial navigation by robustly constraining radar estimates without discarding measurements, enabling more reliable long-term odometry in challenging indoor environments.

Abstract

Ego-velocity estimation from point cloud measurements of a millimeter-wave frequency-modulated continuous wave (mmWave FMCW) radar has become a crucial component of radar-inertial odometry (RIO) systems. Conventional approaches often exhibit poor performance when the number of outliers in the point cloud exceeds that of inliers, which can lead to degraded navigation performance, especially in RIO systems that rely on radar ego-velocity for dead reckoning. In this paper, we propose CREVE, an acceleration-based inequality constraints filter that leverages additional measurements from an inertial measurement unit (IMU) to achieve robust ego-velocity estimations. To further enhance accuracy and robustness against sensor errors, we introduce a practical accelerometer bias estimation method and a parameter adaptation rule that dynamically adjusts constraints based on radar point cloud inliers. Experimental results on two open-source IRS and ColoRadar datasets demonstrate that the proposed method significantly outperforms three state-of-the-art approaches, reducing absolute trajectory error by approximately 36\%, 78\%, and 12\%, respectively.

Figures (7)

• Figure 1: An example of 4D point cloud measurement from an mmWave FMCW radar of the open-source IRS dataset DoerIros2021. The radar's 3D position points are projected onto a 2D image plane (red points), with each number indicating the corresponding 1D Doppler velocity.
• Figure 2: Block diagram overview of the proposed CREVE.
• Figure 3: Our idea visualization using the IRS dataset DoerIros2021, comparing the estimated $\hat{a}_f^r$ calculated from \ref{['a_cal']}, MoCap ground truth, and average acceleration derived from $\hat{v}^r$. For this example, we used $\gamma_{\textrm{min}} = 0.05 \!\!\!\! 0.05 \!\!\!\! 0.05^\top$ m/s and $\gamma_{\textrm{max}} = 1.25 \!\!\!\! 1.25 \!\!\!\! 1.25^\top$ m/s (best viewed in color).
• Figure 4: Relationship between the adaptive parameter $\gamma$ and the inlier ratio $r^2$. These results are drawn from the ColoRadar dataset (arpg_lab_run1 trial), with $\gamma_{\textrm{min}} = 0.04 \!\!\!\! 0.04 \!\!\!\! 0.04^\top$ m/s and $\gamma_{\textrm{max}} = 2 \!\!\!\! 2 \!\!\!\! 2^\top$ m/s.
• Figure 5: Accelerometer bias estimation of our CREVE, with a low-pass filter passband set to be 0.01 Hz.