Frequency Domain Transformation-Aided LiDAR-Inertial Pose Estimation for Autonomous Vehicles Using Implicit Neural Map

Point cloud map-based pose estimation constitutes the cornerstone of autonomous vehicle navigation systems, yet existing techniques suffer from accuracy degradation that depends on resolution. Excessively dense point clouds impose prohibitively high computational loads, whereas overly sparse representations compromise feature distinctiveness, thereby constraining the reliability of real-time positioning. Implicit neural field map emerges as a promising alternative, offering lightweight representation and high-resolution reconstruction. However, current implicit map-based localization methods struggle with initialization, state estimation accuracy, and real-time performance. To address these challenges, this work proposes a novel LiDAR-inertial localization system for vehicles that leverages implicit neural maps augmented by frequency-domain descriptors. A prior map is firstly constructed using neural point models and corresponding descriptors. During the localization process, point cloud frames are converted into Bird's Eye View (BEV) images, from which descriptors are extracted using frequency domain transformation. These descriptors are then used to search the map database and obtain an initial pose estimate. The scan-to-map registration is performed by aligning the point cloud to the implicit neural model. And the short-term high-precision characteristics of the inertial navigation system are utilized to provide state prediction, improving real-time performance. Finally, the state estimation is refined and output through factor graph optimization. Extensive experiments conducted on both public and custom datasets demonstrate that the proposed algorithm outperforms state-ofthe-art methods in terms of accuracy and efficiency