Incorporating Point Uncertainty in Radar SLAM

TL;DR

Radar SLAM is robust to weather but hampered by sparse, noisy radar points. The authors model radar point uncertainty in polar coordinates and integrate it into a radar-inertial odometry framework via uncertainty-aware residuals and a probabilistic data association scheme, including a Doppler residual and point-to-landmark matching. Key contributions are a polar-coordinate uncertainty model for 3D radar measurements, uncertainty-guided weighting in a sliding-window factor graph, and ablations demonstrating performance gains on public and self-collected datasets. The results show improved motion estimation accuracy, underscoring the value of per-point uncertainty for robust radar SLAM in real-world scenarios, with code and data released for reproducibility.

Abstract

Radar SLAM is robust in challenging conditions, such as fog, dust, and smoke, but suffers from the sparsity and noisiness of radar sensing, including speckle noise and multipath effects. This study provides a performance-enhanced radar SLAM system by incorporating point uncertainty. The basic system is a radar-inertial odometry system that leverages velocity-aided radar points and high-frequency inertial measurements. We first propose to model the uncertainty of radar points in polar coordinates by considering the nature of radar sensing. Then, the proposed uncertainty model is integrated into the data association module and incorporated for back-end state estimation. Real-world experiments on both public and self-collected datasets validate the effectiveness of the proposed models and approaches. The findings highlight the potential of incorporating point uncertainty to improve the radar SLAM system. We make the code and collected dataset publicly available at https://github.com/HKUST-Aerial-Robotics/RIO.

Figures (8)

• Figure 1: Factor graph representation of our radar-inertial odometry. Three residuals as factors are constructed to estimate robot poses and map landmarks.
• Figure 2: To obtain the orthonormal basis $(\mathbf{N}_1,\mathbf{N}_2)$ in the angent plane of $p_k \in \mathbb{S}^2$. $R_3(p_k)$ is defined by rotating $e_3$ to $p_k$, $\mathbf{N}_1$and $\mathbf{N}_2$ are obtained through rotating $e_1$ and $e_2$ by $R_3(p_k)$ respectively.
• Figure 3: Given a radar point $p_k$, the matched map landmark is $l_p$ using our probability-guided point matching approach though $l_q$ is closer in Euclidean space. The colors represent the probability density.
• Figure 4: Our customized vehicle for experimental validation. The vehicle is equipped with a radar sensor, an IMU and a computing platform. We validate our proposed approaches within a controlled room outfitted with a motion capture system.
• Figure 5: (a) Visualization of 3-$\sigma$ ellipses with the proposed polar measurement model, and (0, 0, 0) is the coordinate of the radar sensor. The colors represent the probability density. (b) We also present the 3-$\sigma$ noise ellipses with cartesian radar measurement models.