The Coastline as a Structural Constraint: Harnessing Scene Geometry for Autonomous Surface Vessel Localization

Derek R. Benham; Joshua G. Mangelson

cs.ROAug 21, 2026

The Coastline as a Structural Constraint: Harnessing Scene Geometry for Autonomous Surface Vessel Localization

Authors: Derek R. Benham, Joshua G. Mangelson

Organizations: Brigham Young University, Provo, UT 84602 USA

Abstract

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.

Figures & tables

Figure 1: Comparison of terrestrial and maritime sensing environments, demonstrating the lack of traditional information present on water compared to land. Terrestrial environments naturally contain abundant nearby structure distributed throughout the sensor field of view, producing dense observations well suited for conventional localization methods. In contrast, the maritime environment is dominated by open water, with sparse observations of land occurring at greater distances and often on only one side of the vehicle. This disparity highlights the fundamental sensing challenges of maritime localization and motivates the need for domain-specific localization strategies.

Figure 2: Geometric model of the coastal environment. Both the LiDAR and monocular vision pipelines assume a locally planar water surface with the coastline defined as the boundary between the water surface and land. The resulting right-triangle geometry between the vehicle, water surface, and shoreline provides simple geometric constraints for globally referenced localization, where

h_{\text{sensor}}

denotes the sensor height above the water surface,

d_{s}

the distance to shore, and

\alpha

the observed declination angle to the shoreline.

Figure 3: LiDAR localization overview. The pipeline consists of three main stages. For ground plane estimation, the native LiDAR range image is utilized to extract a local water-surface plane, providing absolute estimates of the vehicle’s roll (

\phi

), pitch (

\theta

), and heave (

z

) with respect to the world frame. For cross-view map generation, the raw point cloud is rectified using these attitude estimates and orthographically projected to a 2D occupancy grid to extract a simulated satellite view of the coastline. Finally, for global registration, this simulated view is aligned via ICP to an a priori map of the coastline to resolve the vehicle’s planar (

x,y

) position and heading (

\psi

). Fusing these vertical and planar estimates yields a full

SE(3)

unary observation factor for integration into a fixed-lag smoother factor graph.

Panel of Figure 4 — Figure 4: LiDAR-based water-plane extraction pipeline. (a) Morphologically filtered range image projected onto the corresponding near-infrared image, with the extracted first-return boundary highlighted in green. (b) Surface-boundary extraction in range-image space, showing rejected detections (black) and the filtered boundary signal (green). (c) RANSAC plane fit to the extracted boundary points (green) within the full LiDAR scan (red). The estimated plane is used to derive absolute roll, pitch, and heave measurements relative to the local water surface.

Figure 5: LiDAR fixed-lag factor graph architecture. Consecutive vehicle poses are connected by KISS-ICP odometry factors. Water-Surface factors provide roll, pitch, and height observations at each scan. Every tenth scan, these observations are fused with cross-view shoreline registration measurements to form a globally referenced Coastal Pose factor that constrains the full

SE(3)

state.

Figure 6: Architecture of the visual coastline localization pipeline. The system branches into two processes based on the semantic segmentation of the input image. If the upper water boundary is classified as sky (horizon line), the vehicle’s roll (

\phi

) and pitch (

\theta

) are estimated and a partial attitude measurement is integrated into the local state estimator. Conversely, if the boundary is classified as a coastline, sequential observations are projected and coalesced into a local submap based on the current odometry estimate. This local submap is then aligned against an a priori satellite map via correlative scan matching. Because visual coastline projections lack the fidelity of LiDAR, heading is assumed observable via a magnetometer, restricting the correlative scan match to an exhaustive search over a discretized

XY

translation window around the prior state estimate. This exhaustive search generates an observability-aware factor that provides a strong geometric constraint on cross-shore distance while remaining loosely constrained along the shore. Finally, the resulting coastline registration factor and the relative odometry estimates from the local state estimator are incorporated into the global state estimator.

Figure 7: Visual coastline localization pipeline. (a) Extracted coastline observation (red) overlaid on the input image. (b) Single-image shoreline projection from image space into geospatial coordinates, overlaid on the satellite coastline map (light blue); the vehicle position is shown as a blue triangle. (c) Accumulated shoreline observations (multiple colors) forming a 10-second shoreline submap, with the dominant shoreline trend estimated using PCA (green line). The persistent breaking waves visible along the left side of the image introduce a local distortion in the projected shoreline that remain evident in the extracted PCA line. (d) Correlative scan-matching likelihood generated by registering the PCA shoreline estimate against the satellite coastline map. The resulting likelihood distribution strongly constrains cross-shore position while remaining weakly constrained along the shoreline.

Figure 8: Hierarchical factor graph architecture for the visual localization pipeline. The local fixed-lag smoother estimates the high-rate

SE_{2}(3)\times\text{biases}

state within a local odometry frame using IMU pre-integration together with horizon, magnetometer, and weak altitude constraints. Marginalized relative pose factors connect the local estimator to a sparse global

SE(3)

pose graph, where temporally aligned orientation observations and a weak altitude prior form unified unary pose constraints. Local trajectory estimates are accumulated into shoreline submaps, which are registered against satellite imagery to provide global

x,y

position constraints.

Figure 9: Overhead view of the three dataset trajectories collected in O‘ahu, Hawaii. Imagery and trajectories are rotated vertically to create a concise visual display with a consistent scale. Axes represent relative local distance in kilometers, with start (green circle) and end (red star) points marked for each trajectory (yellow).

Dataset Name	Length (km)	Time (min)	Velocity (m/s)	Mean Dist. Shore (m)	Max Dist. Shore (m)
Makali‘i Point	2.3	18	2.2	25.0	48
Kāne‘ohe Bay	1.5	15	1.8	41.0	67
‘Anahulu River	1.1	16	1.2	7.5	16

Table 1: Dataset characteristics and environmental scale

Dataset	KISS-ICP		FORM		LOAM		LIO-SAM		GenZ-ICP
Dataset	$T$	$R$	$T$	$R$	$T$	$R$	$T$	$R$	$T$	$R$
Makali‘i Point	42.0	2.6	80	11	123	15	844	56	–	–
Kāne‘ohe Bay	22.4	3.0	58	14	1153	120	148	17	61	41
‘Anahulu River	5.1	0.8	4.1	0.8	20	4.2	6.1	1.4	–	–

Table 2: LiDAR odometry comparison. Metrics are presented as translation [m] / rotation [deg] (ATE RMSE). Bold denotes best results, and ‘–’ indicates failure to produce a solution.

Mission Name	Roll Error (deg)	Pitch Error (deg)	Z Error (m)
Makali‘i Point	$-0.11\pm 0.86$	$\phantom{-}0.02\pm 0.67$	$-0.04\pm 0.09$
Kāne‘ohe Bay	$-0.02\pm 0.56$	$-0.18\pm 0.28$	$-0.01\pm 0.06$
‘Anahulu River	$\phantom{-}0.04\pm 1.86$	$-0.61\pm 1.07$	$\phantom{-}0.41\pm 0.39$

Table 3: LiDAR ground plane estimation accuracy across environments

Dataset	Coastal-KISS w/ Satellite			Coastal-KISS (Water Plane)			KISS-ICP
Dataset	$T_{xy}$	$T_{z}$	$R$	$T_{xy}$	$T_{z}$	$R$	$T_{xy}$	$T_{z}$	$R$
Makali‘i Point	7.2	0.14	2.4	24.8	0.13	2.1	40.3	12.2	2.6
Kāne‘ohe Bay	8.1	0.07	3.2	20.4	0.06	2.9	22.4	0.8	3.0
‘Anahulu River	4.5	0.6	3.3	16.5	0.6	5.1	4.8	1.8	0.8

Table 4: Trajectory comparison of the proposed Coastal-KISS framework. Metrics are reported as translation

T_{xy}

T_{z}

[m] and rotation

R

[deg] (ATE RMSE).

Figure 10: Comparison of trajectories produced by the proposed Coastal-KISS pipeline and KISS-ICP baseline across the three evaluation datasets.

Model	mIoU	Water IoU	Sky IoU	Land IoU	Shoreline 95% Error (px)	Shoreline Failure	FPS
Grounded-SAM 2	0.95	0.99	0.90	0.96	7.0	0%	17.5
Grounded-MobileSAM	0.95	0.99	0.90	0.96	5.7	0%	5.2
YOLOE	0.88	0.98	0.72	0.92	9.0	0%	117
YoloWorld-SAM 2	0.64	0.85	0.35	0.73	6.4	13%	6.8
YoloWorld-MobileSAM	0.66	0.86	0.36	0.76	5.2	13%	17.5
SAM 3	0.89	0.97	0.90	0.81	9.3	0%	6.1

Table 5: Comparison of Semantic Segmentation and Shoreline Extraction Performance Across Vision Architectures. Bold and underlined values indicate the best and second-best results, respectively.

Figure 11: Comparison of semantic segmentation models for classifying water, sky, and land. The red curve denotes the extracted upper water boundary used for shoreline localization. Unless otherwise noted, all models are evaluated in a zero-shot setting without task-specific training. Grounded-SAM 2 and Grounded-MobileSAM produce the most consistent semantic segmentations and mask boundaries, although Grounded-SAM 2 struggles in the fourth image where calm-water reflections introduce ambiguity. YOLOE occasionally hallucinates sky within the water or misclassifies portions of the shoreline, most notably labeling the beach as water in the second image. The YOLO-World variants frequently under-segment the sky or misclassify water as land. SAM 3 produces robust segmentations but yields less precise shoreline boundaries and may omit portions of the land. YOLO-seg, trained on 350 labeled images from a single location on the island, fails to generalize to new coastal environments. These results suggest that transformer-based open-vocabulary detectors, such as Grounding DINO, paired with foundation segmentation models such as MobileSAM and SAM 2, provided the most reliable classification across diverse coastal scenes and were therefore selected for use in the proposed localization pipeline.

Makali‘i Point (18.3 min)
	Baseline	Proposed Framework			External Baselines
Metric	Odom Only	Fixed Lag	Hierarchy	Full History	VINS Mono	ORB SLAM3
$T_{\text{err}}$	42.9	13.9	12.1	12.3	–	–
$R_{\text{err}}$	1.4	1.4	1.7	1.4	–	–
Runtime	6.5	15.2	15.7	71.8	–	–
Kāne‘ohe Bay (15.0 min)

Table 6: Visual trajectory accuracy comparison. Metrics are reported as translational error

T_{\mathrm{err}}

[m] and rotational error

R_{\mathrm{err}}

[deg] (ATE RMSE). Runtime [min] includes the complete image-processing and optimization pipeline. – indicates failure to process the entire dataset.

Figure 12: Comparison of trajectories produced by the proposed vision pipeline and visual odometry baselines across the three evaluation datasets.

Dataset	KISS-ICP		FORM		LOAM		LIO-SAM		GenZ-ICP
Dataset	$T$	$R$	$T$	$R$	$T$	$R$	$T$	$R$	$T$	$R$
Makali‘i Point	70.0	5.0	71	13	390	75	$\dagger$	$\dagger$	372	46
Kāne‘ohe Bay	15.3	3.3	$\dagger$	$\dagger$	839	119	$\dagger$	$\dagger$	460	41
‘Anahulu River	2.8	0.5	6.4	1.4	45.6	7.9	$\dagger$	$\dagger$	2.6	0.4

Table 7: LiDAR odometry ablation with semantic land masking constraints. Metrics are presented as translation [m] / rotation [deg] (ATE RMSE). Bold entries denote an improvement over the unmasked baseline (Table 2 ), and

\dagger

marks a regression to failure induced by the land mask.

Figure 13: Absolute translational trajectory error as a function of mission time for all three datasets. Both the proposed LiDAR and monocular localization frameworks substantially outperform conventional LiDAR and visual-inertial methods. Although trajectory error may temporarily increase during extended sections of locally linear coastline, shoreline-to-satellite registration provides intermittent global corrections that prevent unbounded drift. As a result, both proposed approaches maintain bounded localization error throughout each mission.

Robot Localization

Robot localization estimates a robot’s position and orientation from onboard sensors, maps, and observations of its surroundings.

See all 177 papers

Jun 9, 2026

Geometric Coastline Localization using Vision-Language Models

Coastline detection in remotely sensed imagery is commonly formulated as pixel-wise segmentation, even though coastlines used in coastal monitoring are ultimately represented as geometric curves and defined by geomorphic proxies such as vegetation lines, dune toes, or cliff edges. We revisit coastline extraction from a representation perspective and formulate the task as geometric boundary localization, where a thin coastline boundary is localized directly as a curve rather than derived from a segmentation mask. Using the New Zealand Coastal Change Dataset (NZCCD) and LINZ aerial imagery, we develop CoastlineVLM-7B, a vision-language model built on the GeoChat-7B/LLaVA-1.5 architecture that jointly performs coastline presence detection, proxy-type classification, and direct coastline grounding as an ordered polyline. We compare it against U-Net, UNet++, DeepLabV3+, and SegFormer trained using one-pixel-wide coastline masks and evaluate localization using tolerance- and distance-based geometric metrics. On the West Coast test set, U-Net provides stronger local boundary proximity, showing better tolerance scores and lower Chamfer and Modified Average Hausdorff distances, while CoastlineVLM-7B achieves the lowest Hausdorff and Earth Mover's distances, indicating reduced worst-case deviation and stronger global structural correspondence. Ablation studies show that geometric localization is driven primarily by direct coastline-grounding supervision and that GeoChat initialization provides a stronger starting representation than generic LLaVA-1.5 initialization. Zero-shot evaluation on the Australian VCMP dataset shows reduced performance for U-Net and CoastlineVLM-7B, but CoastlineVLM-7B achieves better geometric localization than U-Net. These results show that direct ordered-polyline grounding can be used as an alternative formulation for representing and localizing coastline geometry.

Rafia Malik, Bernhard Pfahringer, Karin Bryan +2

The University of Waikato, New Zealand · The University of Auckland, New Zealand

Remote Sensing Image Understanding Geospatial Foundation Models Large Vision-Language Models Remote Sensing Vision-Language Grounding

Ask PDF

Aug 31, 2026

HorizonNet for visual terrain navigation

This paper investigates the problem of position estimation of unmanned surface vessels (USVs) operating in coastal areas or in the archipelago. We propose a position estimation method where the horizon line is extracted in a 360 degree panoramic image around the USV. We design a CNN architecture to determine an approximate horizon line in the image and implicitly determine the camera orientation (the pitch and roll angles). The panoramic image is warped to compensate for the camera orientation and to generate an image from an approximately level camera. A second CNN architecture is designed to extract the pixelwise horizon line in the warped image. The extracted horizon line is correlated with digital elevation model (DEM) data in the Fourier domain using a MOSSE correlation filter. Finally, we determine the location of the maximum correlation score over the search area to estimate the position of the USV. Comprehensive experiments are performed in a field trial in the archipelago. Our approach provides promising results by achieving position estimates with GPS-level accuracy.

Bertil Grelsson, Andreas Robinson, Michael Felsberg +1

Computer Vision Laboratory, Department of Electrical Engineering, Linköping University, Sweden · Inception Institute of Artificial Intelligence, Abu Dhabi, UAE

Robot Navigation Map-Based Localization Robot Localization Maritime Navigation

Ask PDF

Jun 29, 2026

AUSLUN: A Fixed-Hover UAV--USV System for GNSS-Denied Maritime Search and Navigation

Global navigation satellite system (GNSS) denial can prevent an unmanned surface vehicle (USV) from both finding a distant vessel and maintaining a globally referenced approach. This paper presents AUSLUN (Automatic UAV Search, Localization, and USV Navigation), a fixed-hover aerial-surface system that uses a coastal unmanned aerial vehicle (UAV), which estimates its own pose through visual-inertial odometry (VIO), as a long-range sensing and navigation anchor. The central design shifts sensing motion from UAV translation to a zoom pod and closes the loop through three coupled elements: polygon-aware annular pod scanning, modality-aware bearing-range localization, and target-relative USV guidance with visual-loss recovery. The same gated recursive estimator uses laser range for the non-cooperative target and datalink range for the cooperative USV. Search-planning simulations show that the adaptive yaw bounds reduce scan time and redundant coverage relative to a matched fixed-sector scan, and GPS-referenced field data show that the gated recursive estimator outperforms non-recursive baselines in localization accuracy. An integrated maritime mission further demonstrates the complete search-to-navigation sequence, including a deliberately triggered visual-loss recovery. These results establish the feasibility and operating boundary of fixed-hover UAV assistance for stationary-target approach in coastal GNSS-denied environments. The source code and a video demonstration are publicly available at https://github.com/xirhxq/pod_search and https://youtu.be/S-5RkJs35JI.

Siyuan Yang, Zikai Jia, Hailiang Kuang +4

School of Aerospace Engineering, Beijing Institute of Technology, Beijing 100081, China · Khalifa University Center for Autonomous Robotic Systems (KUCARS), Khalifa University, Abu Dhabi 127788, United Arab Emirates

Robot Navigation GNSS-Denied Localization Cooperative Localization Unmanned Surface Vehicle Maritime Navigation

Ask PDF

Abstract

Robot Localization

Explore similar work

Geometric Coastline Localization using Vision-Language Models

HorizonNet for visual terrain navigation

AUSLUN: A Fixed-Hover UAV--USV System for GNSS-Denied Maritime Search and Navigation