In this paper a novel monocular 3D object detection algorithm, namely MonoStrong, is constructed as a new solution to enhance the navigation safety for general and autonomous ships in particular with the assistance of visual detections. Initially, a channel attention-enhanced feature network is designed to adaptively strengthen feature channels corresponding to critical geometric structures, improving the representation of complex ship characteristics, particularly for feature-sparse ships at medium and long distances. Subsequently, a multi-scale feature fusion module with attention mechanisms is formulated to enhance the robustness of perception through cross-level feature interaction, further strengthening feature representation for ships with scale variations. An optimized 2D-3D bounding box association loss function is then constructed to effectively reduce errors caused by rapid heading variations at short distances and localization drift at extended ranges. A multi-scenario 3D visual dataset for ships, called Kitti3D-Ship, is constructed to validate the proposed technique. The results demonstrate that MonoStrong outperforms existing methods in ship attitude estimation, particularly in longrange, high-density, and multi-scale scenarios. The method offers a reliable solution for intelligent maritime perception and contributes to the safe operation of ships.

The low-altitude economy is rapidly emerging as a new growth engine that integrates aerial infrastructure, ubiquitous connectivity, and intelligent services. Low Earth Orbit (LEO) satellite networks are a representative space-networking infrastructure for this economy, but rapid topology variation, intermittent inter-satellite links, limited onboard resources, and hotspot/burst traffic make load-balanced routing vulnerable to congestion, stale next-hop selection, and packet loss. To address the above issues, this paper proposes a Satellite-Aware Traffic Multi-Agent Load-Balancing (SAT-MALB) scheme, which employs a POMDP-based multi-agent deep reinforcement learning method to solve cooperative traffic distribution, adaptive path selection, and congestion avoidance problems in dynamic LEO networks. First, to improve generalization and cooperative value estimation under non-stationary multi-agent interactions, a Dual-QMIX (DQMIX) hybrid mixing network with TD-error-based soft switching is designed. Next, to enhance routing robustness under partial observability, a GRU-based Deep Q-Recurrent Network (DQRN) is adopted to exploit temporal observation dependencies. Finally, to align routing decisions with reliability and onboard deployment constraints, a PDR-aware objective/reward design and a centralized-training with decentralized-execution (CTDE) mechanism are developed. Extended simulation results show that, under the most challenging 432-satellite burst-hotspot scenario, SAT-MALB reduces the average end-to-end latency by 14.2%, 14.4%, and 8.6% compared with queue-aware Dijkstra, DQN, and QMIX, respectively; it also improves packet delivery ratio by 3.3, 4.0, and 1.6 percentage points.

Intelligent navigation decision support systems are crucial for maritime safety, yet these systems frequently exhibit limited adaptability and reliability in novel, non-predefined scenarios, constituting a persistent challenge. This study proposes Navigation-GPT, a dual-core large language model (LLM) agent designed for intelligent marine navigation. The framework leverages the strong generalization capability of LLMs in unfamiliar situations. It employs a large-scale LLM with ReAct prompting as its control core, responsible for task parsing, planning, and orchestrating external tools to mitigate hallucinations. Furthermore, we fine-tune a lightweight LLM in two stages: using LoRA and a novel rule-controlled GRPO (RC-GRPO) method to develop a specialized agent decision core. This core generates COLREGs-compliant high-level collision avoidance decisions, which are translated into dynamically feasible reference trajectories using a ship dynamics model formulated according to Fossen's equations. A PID-based controller then tracks these trajectories to guide the ship through the resulting avoidance maneuver. Experimental results show that Navigation-GPT completes the process from task reception to decision output in 11.13 s, remaining within the critical safety window for collision avoidance, though longer than the 0.73 s of traditional methods. In complex scenarios, it achieves an 86% collision avoidance success rate and a 90% behavioral compliance rate, outperforming its base model Qwen2.5-7B by 38% and surpassing benchmarks including the dynamic window approach, artificial potential field, and other LLMs (Qwen2.5-0.5B, Qwen2.5-14B, DeepSeek, GPT-4o). This work integrates LLM technology with traditional navigation systems, offering a comprehensive solution that enhances both safety and operational efficiency across diverse maritime scenarios.

It is evident that discrepancies in ship navigation can result in many consequences, harming navigational safety, economic efficiency, and environmental sustainability. Such consequences may include groundings, collisions, augmented fuel consumption, and extended voyage distances. These outcomes are frequently attributable to adverse marine ecological factors, with wind, waves, and currents exerting the most significant influence. To address this issue, the present paper proposes a solution that utilizes the Maneuvering Motion Group (MMG) model as a force-coupling constraint and designs an Embedded Model Predictive Control (E-MPC). This methodology has been demonstrated to reduce forces and moments acting on the vessel in the sway and yaw directions, whilst concomitantly increasing forces in the surge direction. The Fossen three-degree-of-freedom model is employed as the vessel motion model. It is embedded within MPC constraints to perform online rolling optimization of vessel motion, thereby fully leveraging its predictive advantages. In conjunction with linear interpolation, the second-order Nomoto model facilitates the direct conversion of vessel control variables into rudder angle and engine speed. Furthermore, the present study employed measured data from a 28,000 DWT bulk carrier navigating severe sea conditions for comparative simulation.

Abnormal or non-standard operations by ship drivers are one of the major factors leading to water traffic accidents, making the development of a real-time and reliable method for detecting ship driver behavior crucial. This paper proposes a novel approach for multi-person behavior recognition based on the Temporal Shift Module (TSM) algorithm for single-target behavior detection. This algorithm integrates multi-target tracking with single-target behavior recognition methods, achieving simultaneous recognition and detection of multiple drivers within the ship’s bridge. Additionally, a dataset named “SC Action” was created for ship bridge behavior, containing data samples from various ship bridge surveillance videos, including over 1,000 video samples of routine and violation behaviors. Experimental results show that this method, while accurately tracking multiple drivers, achieved a behavior recognition accuracy of 81.72

In the domain of autonomous ship navigation, the construction of bird’s-eye view (BEV) layouts for waterways has obvious significance. A helmsman can generate the BEV layout of the waterway using his/her eyes only. To simulate this intelligence, a novel neural network-based algorithm named Waterway-BEV is proposed, which enables reconstructing a local map formed by the waterway layout and ship occupancies in the bird’s-eye view given a first person view monocular image only. Waterway-BEV employs an efficient SEResNeXt encoder to extract features from first person view (FPV) monocular images, capturing deep semantic information related to waterways and ships. Due to the variations in information across different perspectives, Waterway-BEV incorporates a Cross-View Transformation Module, which takes the constraint of cycle consistency between views into account and makes full use of their correlation to strengthen the view transformation and scene understanding. To fully leverage the feature output of the SEResNeXt encoder, Waterway-BEV employs a decoder based on a dedicated lightweight network. This decoder is responsible for decoding the enhanced bird’s-eye view (BEV) feature maps and generating the BEV layout. By employing the Focal Loss as the loss function for model optimization, Waterway-BEV takes into account the quantity and classification difficulty of ship samples during the training process, thereby improving the generation performance and convergence speed. The experiments demonstrated that Waterway-BEV achieved notable performance metrics, with mIOU and mAP rates reaching 97.8% and 98.2%, respectively, in waterway bird’s-eye view layout generation. Waterway-BEV outperformed other state-of-the-art (SOTA) algorithms in generating BEV layouts of waterways. In particular, during specialized scenarios such as crossroads of waterways and tasks involving small target ships, Waterway-BEV consistently generated satisfactory bird’s-eye view layouts, demonstrating robustness and applicability.

Existing navigation decision support systems often perform poorly when handling non-predefined navigation scenarios. Leveraging the generalization capabilities of large language model (LLM) in handling unknown scenarios, this research proposes a dual-core framework for LLM applications to address this issue. Firstly, through ReAct-based prompt engineering, a larger LLM core decomposes intricate navigation tasks into manageable sub-tasks, which autonomously invoke corresponding external tools to gather relevant information, using this feedback to mitigate the risk of LLM hallucinations. Subsequently, a fine-tuned and compact LLM core, acting like a first-mate is designed to process such information and unstructured external data, then to generates context-aware recommendations, ultimately delivering lookout insights and navigation hints that adhere to the International Regulations for Preventing Collisions at Sea (COLREGs) and other rules. Extensive experiments demonstrate the proposed framework not only excels in traditional ship collision avoidance tasks but also adapts effectively to unstructured, non-predefined, and unpredictable scenarios. A comparative analysis with DeepSeek-R1, GPT-4o and other SOTA models highlights the efficacy and rationality of the proposed framework. This research bridges the gap between conventional navigation systems and LLMs, offering a framework to enhance safety and operational efficiency across diverse navigation applications.

ObjectiveInstance segmentation of ships plays a crucial role in tasks such as monitoring, identification, and tracking, thereby supporting intelligent navigation. However, the wide variability in ship shapes and scales, coupled with environmental interference, leads to poor performance of existing methods perform in Ship Contour extraction. To address this issue, this paper proposes a novel Ship Contour method based on curve recursion. Method By enhancing CenterNet with hierarchical feature extraction and integrating the Deep Layer Aggregation-60 backbone network, the proposed method achieves a balance between accuracy and speed. The Block structure is optimized, and an ECA channel attention mechanism is incorporated to strengthen feature extraction, while the Mish activation function replaces ReLU to improve adaptability in deep learning. In addition, a translation-invariant contour deformation method and a dynamic matching loss function are introduced to accelerate the final contour extraction. Results On the dedicated 2023Ship-seg dataset containing 2300 samples, the proposed method achieved an average precision of AP0.5∶0.95 = 64.0% and a recall rate of AR0.5∶0.95 = 67.9%, outperforming all mainstream instance segmentation algorithms. Conclusion The method can significantly improve visual processing performance in ship monitoring and intelligent navigation scenarios.

Maritime risk identification technology is more and more important for ships sailing on the sea, and can ensure the safety of ships sailing on the sea. However, traditional risk identification methods rely on historical data and qualitative experience, which is difficult to deal with complex and changeable environment. Recently, advancements in artificial intelligence, particularly large language models, have introduced new solutions to these challenges. While these models excel in data processing and pattern recognition, their effective application within the maritime domain remains problematic. This paper presents the ReAct (Reasoning and Acting) model, which enhances the adaptability of large language models by simulating human reasoning and actions. This paper first reviews the traditional methods and existing problems of maritime risk identification, and then discusses the basic principle of ReAct model and its potential in risk identification. Finally, the study proposes the integration of ReAct with large language models, highlighting the advantages and future prospects of this approach. This research aims to provide new theoretical and technological support for maritime risk identification, ultimately contributing to improved risk management and accident prevention.

Ship instance segmentation technology is becoming more and more important in applications such as ship identification, monitoring and tracking, which is of great significance for maritime safety management. However, due to the different shapes of ships, as well as the complexity and volatility of illumination, weather and other factors, the existing ship instance segmentation methods are often difficult to achieve good segmentation results. To address this problem, this paper proposes a real-time segmentation method for ship instances based on contours that uses CenterNet algorithm to detect ship targets. The core network uses DLA-34 (Deep Layer Aggregation) to ensure the detection accuracy and speed. Then, the Deep Snake method is used to segment the ship object accurately. In order to verify the effectiveness of the proposed algorithm, this study constructed a dedicated dataset of 2300 images involving complex environments such as inland rivers and ports under three typical conditions: day, night, and haze. In addition, this dataset was used to test the proposed method, and the average recall (AR) rate was 95.5

Under low-visibility foggy weather, traditional deep-learning-based ship detection methods face challenges, such as poor direct detection performance and the potential introduction of artifacts by defogging processes. To alleviate this problem, we propose a mutual driven ship detection network (MDS-net) with detection and defogging branches in foggy conditions to learn high- and low-level information. Initially, high-level features are employed to drive the defogging branch to generate semantic texture via the shared weights. Subsequently, we share the low-level features of the defogging branch with ship detection by the proposed cross-scale feature fusion (CSFF) module to compensate for the lack of detection feature information. Finally, we introduce a joint optimization strategy to balance the contributions of high-level (ship detection) and low-level (defogging) branches. Experimental results show that our MDS-net outperforms many other state-of-the-art methods in foggy conditions.

ObjectiveTo meet the requirements of remotely controlling ship in curved, narrow and crowded inland waterways, this paper proposes an approach that consists of CNN-based algorithms and knowledge based models under ship-shore cooperation conditions. MethodOn the basis of analyzing the characteristics of ship-shore cooperation, the proposed approach realizes autonomous perception of the environment with visual simulation at the core and navigation decision-making control based on deep reinforcement learning, and finally constructs an artificial intelligence system composed of image deep learning processing, navigation situation cognition, route steady-state control and other functions. Remote control and short-time autonomous navigation of operating ships are realized under inland waterway conditions, and remote control of container ships and ferries is carried out. ResultsThe proposed approach is capable of replacing manual work by remote orders or independent decision-making, as well as realizing independent obstacle avoidance, with a consistent deviation of less than 20 meters. ConclusionsThe developed prototype system carries out the remote control operation demonstration of the above ship types in such waterways as the Changhu Canal Shenzhou line and the Yangtze River, proving that a complete set of algorithms with a CNN and reinforcement learning at the core can independently extract key navigation information, construct obstacle avoidance and control awareness, and lay the foundation for inland river intelligent navigation systems.