Answers are grounded in this paper and cite their sources.
Unmanned surface vehicles (USVs) require stable three-dimensional maps for port inspection, nearshore monitoring, and autonomous berthing. In dynamic water-surface environments, however, moving vessels and reflection-induced LiDAR returns contaminate front-end point clouds. These observations weaken registration constraints, introduce map artifacts, and obscure static structures near the water–shore interface. We propose a cross-modal semantic–geometric filtering module for USV LiDAR simultaneous localization and mapping (SLAM). The module converts visual semantic probabilities into point-level LiDAR confidence through temporal association, camera–LiDAR calibration, and continuous pixel-wise probability querying. This confidence provides a soft prior for interference regions. A water-surface boundary distance constraint and a vessel-range-adaptive attenuation mechanism then selectively suppress reflection-induced anomalous points and residual dynamic-vessel observations while preserving valid waterside structures. In the primary quantitative port-basin sequence S1, Joint reduced point counts by 87.18% in the vessel-residual region and 77.98% in the reflection-affected water region, while retaining 91.37% in the building region. Sim(3)-aligned absolute pose error RMSE decreased from 2.591 m to 1.962 m (24.28%). Regional statistics, spatial map comparisons and protection-term ablations provide complementary evidence that joint semantic–geometric selection balances interference suppression and structural preservation, with improved aligned trajectory consistency under the evaluated conditions.
No takes yet. Share an insight, caveat, or question.
Song et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: