RaLL: End-to-End Radar Localization on Lidar Map Using Differentiable Measurement Model

Compared to the onboard camera and laser scanner, radar sensor provides lighting weather invariant sensing, which is naturally suitable for long-term localization under adverse conditions. However, data sparse noisy, resulting in challenges mapping. On other hand, most popular available map currently built by lidar. In this paper, we propose an end-to-end deep learning framework Radar Localization on Lidar Map (RaLL) bridge gap, not only achieves robust but also exploits mature lidar mapping technique, thus reducing cost of We first embed both modals into a common feature space neural network. Then multiple offsets are added modal exhaustive similarity evaluation against current modal, yielding regression pose. Finally, apply differentiable measurement model Kalman Filter (KF) learn whole sequential process manner. \textit{The system with network based at front-end KF back-end.} To validate feasibility effectiveness, employ multi-session multi-scene datasets collected from real world, results demonstrate that our proposed superior performance over $90km$ driving, even generalization scenarios where training UK, while testing South Korea. release source code publicly.