Source-linked AI summary
The Coastline as a Structural Constraint: Harnessing Scene Geometry for Autonomous Surface Vessel Localization
Derek R. Benham, Joshua G. Mangelson
TL;DR
GPS-denied ASV localization is difficult because conventional systems drift without global position and coastal scenes offer sparse traditional structure. The paper uses shoreline and water-surface geometry in complementary LiDAR and monocular pipelines, finding improved LiDAR trajectories and bounded visual drift across three coastal datasets.
Problem
GPS-dependent systems drift without regular global position, while conventional SLAM lacks an external global reference for maritime operation.
Method
The paper combines LiDAR water-surface estimation and shoreline-to-satellite registration with monocular semantic shoreline and horizon extraction, submaps, and hierarchical factor-graph fusion.
Results
Across three real-world coastal datasets, Coastal-KISS consistently improved trajectory accuracy while the visual framework achieved globally bounded localization through shoreline-to-satellite registration.
Takeaways & Limitations
Coastal geometry provides a powerful source of globally referenced information for GPS-denied maritime localization, and zero-shot segmentation supports shoreline extraction across diverse environments.
Takeaways & Limitations
The vision frontend can fail around vessels, and tidal fluctuations may degrade alignment between observed shorelines and satellite maps.
Abstract
from arXiv · showhide
Coastal environments contain rich, largely unexploited geometric structure capable of providing globally referenced localization cues. In this work, we present two complementary localization frameworks that exploit shoreline and water-surface geometry for GPS-denied autonomous surface vessel localization. The first framework leverages LiDAR observations of the water surface to estimate roll, pitch, and heave (vertical motion), while recovering global position and heading through direct registration of shoreline observations against a satellite-derived coastline map. The second framework relies solely on passive imagery to detect the shoreline and horizon through semantic segmentation. Using the proposed coastal scene geometry, shoreline distance is inferred from monocular imagery. Shoreline observations are accumulated into short-duration local submaps, registered against the same satellite-derived coastline map, and fused within a hierarchical factor graph. Evaluated across three real-world coastal datasets, the LiDAR pipeline consistently improves trajectory accuracy over standard baselines, while the monocular architecture maintains bounded long-term drift. In addition, we establish that modern zero-shot foundation models can reliably extract shoreline observations across diverse coastal environments. Together, these results demonstrate that coastal geometry provides a powerful and dependable source of globally referenced information for GPS-denied maritime localization.
I. INTRODUCTION
Autonomous surface vessels need robust localization because GPS-dependent systems drift during outages, while conventional mapping and SLAM approaches face global-reference and computational limitations. This paper exploits persistent coastal geometry through complementary LiDAR and monocular pipelines, evaluated across real-world coastal datasets.
- GPS-dependent ASV systems ultimately drift and fail without regular global position estimates.
- Prior-map localization restricts operation to mapped regions and depends on reference-map fidelity and recency.
- Traditional SLAM provides only a local frame without external global references, while long missions can make graph optimization computationally prohibitive.
- Coastal scenes provide a reference plane, horizon, and persistent shoreline observable from vehicle sensors and satellite imagery.
- The proposed LiDAR and monocular pipelines exploit water-surface and shoreline geometry for globally referenced localization.
- The work evaluates the unified framework across three real-world coastal datasets totaling nearly 5 km of operation.
II. RELATED WORK
Related work shows that maritime localization methods often trade global reference, robustness, and environmental coverage against sensing and computational constraints. Existing LiDAR, radar, and vision systems motivate specialized approaches that exploit coastal structure directly.
- Early ASV state-estimation methods relied on GPS or operated in sheltered waters where wave-induced motion was negligible.
- Radar: Radar is robust to water clutter and atmospheric obscurants but provides coarser geometry because of lower angular resolution and multipath interference.
- LiDAR: LiDAR offers denser, more precise observations than radar but has reduced range and greater sensitivity to water-surface reflections and backscatter.
- LiDAR: Existing LiDAR maritime methods improved odometry robustness but remained fundamentally dependent on GNSS for global position.
- Cameras: Vision-based odometry and SLAM methods remain constrained by shoreline proximity or fail to provide globally referenced coordinates in open-water settings.
B. Cross-View Localization
Cross-view localization aligns vehicle observations with geo-referenced overhead maps to bound drift, but prior methods often lacked temporal consistency, broad modality coverage, or full coastal scene structure. Semantic segmentation and zero-shot foundation models provide a route to extracting the coastal boundaries needed by the proposed framework.
- Cross-view localization recovers global position by registering local maps against geo-referenced overhead representations such as satellite imagery.
- Independent radar-scan alignment can outperform dead reckoning but limits smooth continuous trajectories between observations.
- Prior satellite-based methods demonstrated bounded drift but remained limited to specific sensing modalities and underused coastal geometric and semantic structure.
- Semantic segmentation can support state estimation using high-level scene understanding while requiring only broad coastal classes for this framework.
- Semantic Coastal Scene Understanding: Traditional maritime vision methods provide task-specific boundaries but lack contextual understanding of whether boundaries represent horizons or coastlines.
- Semantic Coastal Scene Understanding: The proposed frontend combines Grounding DINO with SAM 2 to extract water, sky, and land observations without task-specific training data.
III. METHODS
The methods derive two localization pipelines from a shared coastal scene model in which the mean water surface and coastline provide stationary geometric references. LiDAR and monocular sensing use these constraints through complementary pose and shoreline-estimation strategies.
- The paper proposes LiDAR and vision methods that derive drift-free geometric constraints from the shoreline and water surface.
- Both pipelines use a common world and vehicle frame model with the mean water surface represented as a planar reference.
- The model assumes time-averaged water-surface elevation remains constant during a mission despite instantaneous wave perturbations.
- The coastline is modeled as the intersection of terrain and mean water surface, with tidal variation assumed negligible over 15–20-minute missions.
B. LiDAR based Localization
The LiDAR pipeline uses the water surface to estimate vertical and attitude states, then registers shoreline geometry against a satellite-derived coastline map to recover planar pose and globally referenced localization.
- LiDAR-Based Ground Plane Estimation: Water-surface plane estimation provides absolute roll, pitch, and heave estimates relative to the world frame.The method extracts the plane from LiDAR range-image returns and fits it with RANSAC.
- LiDAR Cross-View Satellite Image Alignment: Rectified shoreline points are orthographically projected into a 2D occupancy grid that approximates a satellite view for cross-view registration.The projection reduces the vertical coordinate after correcting the point cloud using estimated roll and pitch.
- LiDAR-Based Ground Plane Estimation: The LiDAR range image is processed structurally to isolate first-return boundaries associated with the shoreline, wake, or wave crests.Morphological closing suppresses isolated returns before a bottom-up column search extracts the boundary signal.
- LiDAR Cross-View Satellite Image Alignment: Localized ICP aligns the simulated satellite view with a prior coastline map while restricting optimization to planar position (x, y) and heading (ψ).Initialization from the current pose estimate limits the search window and helps prevent convergence to local minima.
- Factor Graph SLAM with LiDAR Odometry: A fixed-lag factor graph fuses KISS-ICP odometry with water-surface and shoreline observations to produce globally referenced pose estimates.Water-plane measurements constrain roll, pitch, and height, while shoreline registration constrains planar position and yaw.
C. Vision-Based Localization
The vision pipeline extracts horizon and shoreline geometry from monocular images, reconstructs shoreline observations, and registers short-duration submaps against satellite coastline maps. A hierarchical factor graph combines local high-rate estimation with sparse global optimization, while shoreline geometry remains less observable along straight coastlines.
- The pipeline uses semantic segmentation, horizon-based attitude estimation, shoreline reconstruction, satellite registration, and hierarchical global estimation.
- Zero-shot Grounding DINO and SAM 2 partition images into water, sky, land, and background classes without task-specific training data.Grounding DINO initializes prompts, while SAM 2 tracks regions across sequential frames.
- Columnwise boundary extraction identifies coastline and horizon points; RANSAC filters horizon points before roll and pitch estimation.
- Shoreline pixels are projected into the world frame and intersected with the mean water surface using known camera height and estimated attitude.
- Sequential shoreline observations are accumulated into local submaps, filtered in a 1 m grid, and reduced with PCA before satellite registration.A 10-second accumulation window supports a locally linear shoreline representation.
- Registration strongly constrains distance perpendicular to the shoreline but remains weak along it, creating longitudinal drift on straight coastline sections.A fixed-lag local smoother cannot retroactively update marginalized states, whereas a full-history graph is computationally intractable over long runs.
IV. RESULTS AND INSIGHTS
The evaluation spans three real-world coastal datasets and assesses both localization frameworks, including the visual perception front-end and a LiDAR odometry ablation. It also examines why conventional LiDAR odometry struggles in coastal environments.
- Three real-world coastal datasets are used to evaluate the LiDAR and monocular localization pipelines independently.
- The vision evaluation additionally examines the perception front-end and includes an ablation study of conventional LiDAR odometry difficulties in coastal environments.
A. System and Dataset Overview
The study evaluates coastal localization using a sensorized autonomous surface vessel across diverse O‘ahu datasets, with RTK-GPS providing quantitative reference measurements. Results show that standard LiDAR odometry often struggles in open water, while coastal geometric constraints improve trajectory estimation.
- A. System and Dataset Overview: The platform combines three global-shutter cameras, an Ouster OS1-128 LiDAR, an SBG Ellipse-D IMU, and dual-antenna RTK-GPS for evaluation.Sensors and onboard computers were synchronized using Precision Time Protocol.
- A. System and Dataset Overview: The reference trajectory uses a fixed-lag factor graph disciplined by dual-antenna RTK-GPS, with RTK-GPS pose nodes added at 5 Hz.
- A. System and Dataset Overview: Three O‘ahu coastal datasets span varied shoreline structures, vegetation densities, stand-off distances, and sea states from calm inland waters to exposed coastal conditions.
- B. LiDAR Results: Most evaluated LiDAR odometry pipelines struggled in open-water datasets, with several failing to produce valid solutions.The reported degradation is associated with absent stable ground planes, complex water returns, wave-induced motion, and sparse persistent returns.
- B. LiDAR Results: On two open-water datasets, water-plane estimation produced roll and pitch mean errors below 0.2° and standard deviations below 1°.Recovered heave remained accurate within 4 cm on average, with less than 1 cm standard deviation.
- B. LiDAR Results: Shoreline-to-satellite registration significantly reduced horizontal drift, producing the most accurate overall trajectories in open-water datasets and matching the best baseline in the river environment.
C. Semantic Segmentation Evaluation
The evaluation compares semantic segmentation architectures for extracting shoreline boundaries from coastal imagery. Zero-shot foundation-model pipelines generally provide more reliable boundaries, while Grounded-MobileSAM offers the strongest accuracy-efficiency trade-off despite prompting overhead.
- Evaluation setup: The evaluation uses region-based IoU, per-class IoU, 95th-percentile shoreline error, and failure rate to assess segmentation and shoreline extraction.Shoreline-specific error is emphasized because registration depends on water-land boundary alignment.
- Generalization: YOLO11-Seg performs well within its training environment but fails to generalize to unseen shorelines and environmental conditions.The model was fine-tuned on 350 labeled images from a small subsection of one dataset.
- Model comparison: Zero-shot foundation-model pipelines provide more reliable shoreline boundaries than YOLO-based approaches across diverse coastal scenes.Grounding DINO paired with SAM 2 or MobileSAM was selected for the proposed localization pipeline.
- Model trade-offs: SAM 3 produces coherent segmentations across varied scenes but has less precise shoreline boundaries and higher computational cost than stronger boundary-focused alternatives.Its boundary imprecision increases localization error despite competitive mean IoU scores.
- Quantitative results: Grounded-MobileSAM achieves the highest overall segmentation accuracy, with the highest mean IoU and lowest shoreline segmentation error.Its reliance on Grounding DINO prompts for every frame limits processing speed.
D. Vision-Based Localization Results
The proposed vision localization framework is evaluated against local, global, visual-inertial, and LiDAR odometry baselines across three coastal datasets. Shoreline submap alignment reduces translational error, while hierarchical global optimization extends corrections beyond the local estimator’s retained history.
- Baseline comparison: Conventional visual baselines fail or drift substantially when nearby land features are limited, including 827 m translational error for VINS-Mono.Neither VINS-Mono nor ORB-SLAM3 successfully processes Makali‘i Point; ORB-SLAM3 also fails in K¯ane‘ohe Bay.
- Vision-based localization results: Shoreline submap alignment substantially reduces translational error relative to the local odometry baseline across all three datasets.Adding the hierarchical global graph yields further reductions, particularly in K¯ane‘ohe Bay and ‘Anahulu River.
- Hierarchical optimization: The hierarchical global graph propagates delayed shoreline corrections beyond the limited history retained by the fixed-lag smoother.This addresses locally linear coastline segments where trajectory error can temporarily increase.
- LiDAR ablation: Restricting LiDAR returns to land improves odometry in the ‘Anahulu River dataset when calm water and abundant land returns are available.The result suggests that removing water-surface reflection returns can improve trajectory estimation under those conditions.
- LiDAR ablation: Open-water datasets show little benefit from land-only LiDAR masking because shoreline returns are sparse, distant, and available from limited viewing directions.The loss of valid land measurements can outweigh gains from suppressing water-surface artifacts.
V. DISCUSSION AND LESSONS LEARNED
The experiments show that exploiting coastal scene geometry enables globally consistent localization in GPS-denied maritime environments.
- Discussion: Exploiting coastal scene geometry enables globally consistent localization in GPS-denied maritime environments.The discussion frames this result alongside sensing, environmental, and architectural trade-offs observed during field evaluation.
A. Maritime Environments Are Not Homogeneous
Coastal environments differ substantially in sensing conditions and geometric observability, producing distinct perception and localization failure modes. These differences create a trade-off between LiDAR’s spatial fidelity and vision’s attitude constraints.
- Maritime Environments Are Not Homogeneous: Calm rivers, sheltered bays, and exposed coastlines provide substantially different quantities and qualities of geometric observations.The three datasets were collected in one geographic region but still exhibited distinct sensing conditions and failure modes.
- Maritime Environments Are Not Homogeneous: Specular water reflections can corrupt LiDAR water-plane estimates and create false shoreline boundaries in the visual pipeline.The resulting errors affect roll, pitch, height, and estimated land distance.
- Maritime Environments Are Not Homogeneous: Conventional LiDAR odometry performs well in the river because dense bank vegetation provides features on both sides, but routinely fails in open water.Open-water scenes offer sparse, distant land returns often available on only one side of the vessel.
- Maritime Environments Are Not Homogeneous: LiDAR directly measures range for high-fidelity shoreline reconstruction and registration, whereas vision estimates shoreline distance from segmentation, calibration, and attitude.The visual pipeline is therefore more sensitive to segmentation and attitude-estimation errors.
- Maritime Environments Are Not Homogeneous: Vision provides informative horizon-based attitude constraints, while LiDAR provides superior spatial fidelity and consistently lower trajectory error.The visual framework depends more on accumulated observations and delayed global corrections.
C. Observability Degradation during Prolonged
The visual localization framework remains vulnerable when coastal geometry provides weak motion observability or when segmentation and map assumptions fail. These limitations emphasize that maritime localization depends on environment-specific observability and assumptions.
- Straightaways: Localization error grows most rapidly along locally linear coastlines because shoreline registration provides little information about motion parallel to the coast.Position error can accumulate in that direction during extended passes.
- Straightaways: Hierarchical smoothing retroactively distributes corrections, but pronounced error accumulation persists during prolonged straightaways when dead reckoning is weak.The paper identifies affected trajectory segments in Makali´i Point and K¯ane´ohe Bay.
- Segmentation and map dependencies: The visual frontend assumes every pixel not classified as water or sky is land, an assumption that breaks down around dynamic or stationary vessels.This introduces vulnerability through the segmentation frontend.
- Broader implications: The results show that maritime localization depends on sensing conditions, observability characteristics, and environmental assumptions rather than sensing capability alone.Different coastal settings introduce distinct sources of uncertainty and failure.