RACIF-SLAM: Reliability-Aware Conflict-Gated LiDAR and W-Band Radar Fusion for Continuous Near-Coastal Localization and Hierarchical Occupancy Mapping

Abstract

Reliable near-coastal localization is difficult because LiDAR and scanning radar fail in complementary regimes: LiDAR provides precise local geometry but loses support over open water, whereas radar preserves long-range observations but suffers from clutter, multipath, and low spatial resolution. We present RACIF-SLAM, a reliability-aware framework that unifies continuous localization and hierarchical occupancy mapping. RACIF-SLAM transforms sensor-native motion increments into a common vessel frame, infers temporally filtered reliability from registration evidence, and applies an innovation gate to suppress the less credible modality under cross-sensor conflict before fusion on SE(2). When LiDAR tracking collapses, a radar bridge sustains motion estimation and re-anchors the recovered LiDAR segment without trajectory discontinuity. For mapping, LiDAR retains local occupied/free-space authority, while spatially broadened radar evidence extends support beyond the effective LiDAR range. In the four evaluated near-coastal segments, every reported pose was supported by at least one accepted sensor increment; the dead-reckoning fallback was not invoked. Against synchronized dual-antenna GNSS, RACIF-SLAM achieves a mean ATE of 0.700 m and a 10-frame translational RPE of 0.183 m, reductions of 66.3% and 65.1%, respectively, over the next-best evaluated method. These segment-specific results indicate that evidence-dependent authority transfer can improve localization continuity while retaining fine local map structure.