FD-RIO: Fast Dense Radar Inertial Odometry

TL;DR

FD-RIO addresses ego-motion estimation under challenging sensing conditions by fusing high-rate IMU measurements with dense radar scans through a Kalman filter. The core approach combines a phase-correlation-based dense radar odometry pipeline (two-step rotation then translation) with a 5-DOF SE(2)-oriented state estimator that asynchronously integrates IMU data and radar-derived motions. The key contributions are the first Kalman-filter fusion of scanning radar with IMU, a compact and real-time dense radar odometry pipeline, and comprehensive evaluation on MulRan and Boreas with ablations and runtime analysis. The practical impact lies in delivering robust, real-time odometry for mobile platforms operating in adverse weather and lighting, with favorable comparisons to state-of-the-art methods and clear pathways for efficiency improvements.

Abstract

Radar-based odometry is a popular solution for ego-motion estimation in conditions where other exteroceptive sensors may degrade, whether due to poor lighting or challenging weather conditions; however, scanning radars have the downside of relatively lower sampling rate and spatial resolution. In this work, we present FD-RIO, a method to alleviate this problem by fusing noisy, drift-prone, but high-frequency IMU data with dense radar scans. To the best of our knowledge, this is the first attempt to fuse dense scanning radar odometry with IMU using a Kalman filter. We evaluate our methods using two publicly available datasets and report accuracies using standard KITTI evaluation metrics, in addition to ablation tests and runtime analysis. Our phase correlation -based approach is compact, intuitive, and is designed to be a practical solution deployable on a realistic hardware setup of a mobile platform. Despite its simplicity, FD-RIO is on par with other state-of-the-art methods and outperforms in some test sequences.

Figures (7)

• Figure 1: Overview of FD-RIO main components. Radar scans are processed by the radar odometry pipeline, its estimates are fed to a Kalman filter. IMU measurements are also sent to the filter where state estimates from both sensors are fused.
• Figure 2: FD-RIO's dense radar odometry pipeline. Translation and rotation between two scans are estimated using a two-step approach: magnitudes of FFT images are warped to polar coordinates, then the phase correlation is calculated to find the rotation, one scan is then corrected for rotation and phase correlation is used again to estimate the translation.
• Figure 3: Two radar scans taken from the Boreas sequence 2021-04-08-12-44. This example illustrates the registration performed by FD-RIO's odometry pipeline despite the noticeable motion distortion and ghost detections. It is important to note that the bright areas are not detections/features; FD-RIO is a dense method that does not rely on detections. The bright areas are stronger returns displayed as pixels with higher intensities for better visualization. (a) Sample frame #1. (b) Sample frame #2. (c) Both frames overlaid highlighting the pose change. (d) Both frames after registration.
• Figure 4: Test results using MulRan dataset. Estimated trajectories by FD-RIO vs ground truth. Sequences shown are: (a) DCC01. (b) DCC02. (c) DCC03. (d) Riverside01. (e) Riverside02. (f) Riverside03. (g) KAIST02. (h) KAIST03. (i) Sejong02. (j) Sejong03
• Figure 5: Test results using Boreas dataset. Estimated trajectories by FD-RIO vs ground truth. Sequences shown are: (a) 2020-11-26-13-58. (b) 2021-01-26-11-22. (c) 2021-04-08-12-44. (d) 2021-04-29-15-55.