RaSCL: Radar to Satellite Crossview Localization

TL;DR

The paper presents a GNSS-free global localization framework that fuses ground mmWave radar with overhead RGB satellite imagery by jointly optimizing radar odometry and satellite-based registration. A learned occupancy predictor converts satellite imagery into a point-cloud-like representation, enabling ICP-based alignment with radar scans, while a factor-graph backbone fuses this global registration with local radar odometry over a sliding window. The method is evaluated on urban, suburban, and marine datasets (Boreas, Oxford, and Boat), showing improved localization over prior radar-to-overhead approaches in several settings and demonstrating robustness to occlusions and unseen terrain. This work advances GPS-independent localization for diverse autonomous platforms, with practical implications for field robotics in unmapped or challenging environments.

Abstract

GNSS is unreliable, inaccurate, and insufficient in many real-time autonomous field applications. In this work, we present a GNSS-free global localization solution that contains a method of registering imaging radar on the ground with overhead RGB imagery, with joint optimization of relative poses from odometry and global poses from our overhead registration. Previous works have used various combinations of ground sensors and overhead imagery, and different feature extraction and matching methods. These include various handcrafted and deep-learning-based methods for extracting features from overhead imagery. Our work presents insights on extracting essential features from RGB overhead images for effective global localization against overhead imagery using only ground radar and a single georeferenced initial guess. We motivate our method by evaluating it on datasets in diverse geographic conditions and robotic platforms, including on an Unmanned Surface Vessel (USV) as well as urban and suburban driving datasets.

Figures (10)

• Figure 1: Top shows our Otter USV data collection platform equipped with a Navtech RAS6 radar, an Ouster OS1 lidar and a NovAtel RTK GPS system. Top right illustrates a radar observation and extracted real radar points in red. Bottom left shows the extracted real radar points and our predicted blue points overlaid onto the satellite image against which we are localizing. Bottom right shows the route driven in yellow.
• Figure 2: Overview of localization pipeline. The red factors are binary factors formed using the solution from point-to-point ICP between consecutive real radar scans. The blue factors are formed by using the most recent state estimate to retrieve a satellite image, indicated by $\boldsymbol{\mathbf{T}} _{k-1}$, then extracting a predicted point cloud (blue) and computing a relative pose registration $\tilde{ \boldsymbol{\mathbf{T}} }_k$ from the live radar scan to the satellite image to form a global unary factor.
• Figure 3: Overview of training pipeline for the learned occupancy representation where $\odot$ is an element-wise product. The network predictions are masked using $\text{M}$ to prevent gradients from flowing through predictions for which we are uncertain. The method for obtaining $\text{M}$ and $\text{I}^{L\dagger}$ is described in Section III-A.
• Figure 4: Odometry drift visualization. In this comparison, we show our optimization without the satellite localization factor (orange) and with (blue). The figure shows our satellite localization factor significantly improves performance in unseen areas and does not require highly accurate odometry. The degraded boat odometry is discussed in Section IV-C.
• Figure 5: Dataset geographic splits. Blue is training, red is test, purple is the validation area. Path length is indicated for each dataset.