RINO: Accurate, Robust Radar-Inertial Odometry with Non-Iterative Estimation

TL;DR

RINO tackles radar–inertial odometry under adverse weather by introducing a non-iterative, uncertainty-aware pose estimation scheme with adaptive voting and dynamic fusion weights. Building on ORORA, it adds motion distortion compensation and explicit propagation of radar pose uncertainty into an adaptive MAP/ESKF fusion framework, yielding improved translation and rotation accuracy across diverse datasets. It demonstrates robust real-time performance on MulRan and Boreas, with real-world validation showing reliability under challenging conditions. This approach enhances autonomous navigation by delivering more robust, sensor-aware localization in environments where vision/LiDAR fail.

Abstract

Odometry in adverse weather conditions, such as fog, rain, and snow, presents significant challenges, as traditional vision and LiDAR-based methods often suffer from degraded performance. Radar-Inertial Odometry (RIO) has emerged as a promising solution due to its resilience in such environments. In this paper, we present RINO, a non-iterative RIO framework implemented in an adaptively loosely coupled manner. Building upon ORORA as the baseline for radar odometry, RINO introduces several key advancements, including improvements in keypoint extraction, motion distortion compensation, and pose estimation via an adaptive voting mechanism. This voting strategy facilitates efficient polynomial-time optimization while simultaneously quantifying the uncertainty in the radar module's pose estimation. The estimated uncertainty is subsequently integrated into the maximum a posteriori (MAP) estimation within a Kalman filter framework. Unlike prior loosely coupled odometry systems, RINO not only retains the global and robust registration capabilities of the radar component but also dynamically accounts for the real-time operational state of each sensor during fusion. Experimental results conducted on publicly available datasets demonstrate that RINO reduces translation and rotation errors by 1.06% and 0.09°/100m, respectively, when compared to the baseline method, thus significantly enhancing its accuracy. Furthermore, RINO achieves performance comparable to state-of-the-art methods.

Figures (16)

• Figure 1: Visualization of the trajectory obtained by RINO on the KAIST02 sequence of the Mulran dataset. The raw radar data and IMU data of the KAIST02 sequence are given in the format of ROS messages using the File player provided by the Mulran dataset and can run robustly in real time after being processed by RINO. The trajectory of the second lap closely overlaps with that of the first lap, demonstrating the effectiveness of RINO.
• Figure 2: Overview of RINO. The IMU branch is responsible for constructing the motion equations and performing motion distortion compensation. The Radar branch extracts and matches keypoints from paired raw radar data. After MCIS, the Radar branch non-iteratively estimates the rotation and translation between two scans, along with their uncertainties. These estimates are then fused with the IMU's uncertainty input to the ESKF update step, leading to the optimal pose estimation through MAP.
• Figure 3: Visualization of the keypoint extraction results. The blue points represent the extracted keypoints, with the background displaying the Cartesian radar tensor. The grayscale values in the radar image correspond to the power of the radar echoes. (a) illustrates the results of the improved cen2018 method; (b) shows the results of the cen2019 method. It is clearly observable that the keypoints extracted by the improved cen2018 are better aligned with the brighter regions in the radar image.
• Figure 4: Visualization of motion distortion compensation results. The blue points represent the extracted keypoints, while the red points indicate the keypoints after motion distortion compensation. (a) shows the result with only Doppler shift compensation; (b) shows the result with both Doppler shift and motion distortion compensation. Note the zoomed-in region on the radar image, where it can be inferred that the red points correspond to keypoints extracted from a wall-like object. It is observable that before distortion compensation, the red points were split into two segments, while after compensation, the red points were restored to nearly a straight line.
• Figure 5: Uncertainty model of rotation and translation estimates. (a) The raw data of scanning radar clearly exhibits significantly higher resolution in the radial direction compared to the angular direction. This results in severe anisotropic uncertainty in the keypoints extracted from it. (b) The variances of the rotation estimates (left) and translation estimates (right) are calculated from the uncertainty model of the keypoints.