4D Radar-Inertial Odometry based on Gaussian Modeling and Multi-Hypothesis Scan Matching

TL;DR

This work addresses robust radar-based odometry under sparse, noisy 4D radar scans by adopting a global, jointly-optimized 3D Gaussian representation for radar geometry, inspired by 3D Gaussian Splatting. It introduces a multi-hypothesis scan matching scheme to protect against local optima and integrates these components into a Radar-Inertial Odometry pipeline using an EKF that fuses inertial data, egovelocity, and scan-matching observations. The approach yields performance comparable to, and in several cases better than, established registration methods across multiple datasets and radar types, while offering improved flexibility and potential for real-time extensions. Practical impact includes more robust radar-based localization in adverse weather, enabling reliable SLAM in conditions where LiDAR or vision-based systems struggle.

Abstract

4D millimeter-wave (mmWave) radars are sensors that provide robustness against adverse weather conditions (rain, snow, fog, etc.), and as such they are increasingly used for odometry and SLAM (Simultaneous Location and Mapping). However, the noisy and sparse nature of the returned scan data proves to be a challenging obstacle for existing registration algorithms, especially those originally intended for more accurate sensors such as LiDAR. Following the success of 3D Gaussian Splatting for vision, in this paper we propose a summarized representation for radar scenes based on global simultaneous optimization of 3D Gaussians as opposed to voxel-based approaches, and leveraging its inherent Probability Density Function (PDF) for registration. Moreover, we propose tackling the problem of radar noise entirely within the scan matching process by optimizing multiple registration hypotheses for better protection against local optima of the PDF. Finally, following existing practice we implement an Extended Kalman Filter-based Radar-Inertial Odometry pipeline in order to evaluate the effectiveness of our system. Experiments using publicly available 4D radar datasets show that our Gaussian approach is comparable to existing registration algorithms, outperforming them in several sequences.

Figures (2)

• Figure 1: Overview of the Radar-Inertial Odometry pipeline including our proposed Gaussian modeling and scan matching processes. Bold font indicates input/output, the highlighted area indicates our contribution
• Figure 2: Qualitative comparison between NDT and our Gaussian modeling using NTU4DRadLM's cp sequence.