PaperlayerView exact revision on arXiv

Source-linked AI summary

Exploiting AIS Data for Intelligent Maritime Navigation: A Comprehensive Survey

Enmei Tu, Guanghao Zhang, Lily Rachmawati, Eshan Rajabally, Guang-Bin Huang

arXiv:1606.00981v1cs.OH

TL;DR

Maritime navigation needs safer and more efficient use of vessel-movement data amid human-error risks and complex traffic conditions. This paper surveys AIS data sources and AIS-based methods for anomaly detection, route estimation, collision prediction, and path planning. It concludes that methods offer complementary capabilities but generally do not detect every anomaly type, while future collision assessment should learn vessel motion from historical AIS trajectories rather than assume constant motion.

  • Problem

    AIS data offer useful safety information, but extracting knowledge is challenged by noisy, incomplete, and high-volume data and by diverse navigation-analysis needs.

  • Method

    The paper surveys AIS data sources and representative methods for traffic anomaly detection, route estimation, collision prediction, and path planning.

  • Results

    Surveyed methods generally detect one or two anomaly types rather than all types, while collision assessment commonly uses constant speed and course at detection time.

  • Takeaways & Limitations

    The paper suggests learning ship motion features from large historical AIS data to predict future paths for more meaningful collision-risk assessment.

Abstract

from arXiv · show

The Automatic Identification System (AIS) tracks vessel movement by means of electronic exchange of navigation data between vessels, with onboard transceiver, terrestrial and/or satellite base stations. The gathered data contains a wealth of information useful for maritime safety, security and efficiency. This paper surveys AIS data sources and relevant aspects of navigation in which such data is or could be exploited for safety of seafaring, namely traffic anomaly detection, route estimation, collision prediction and path planning.

I. INTRODUCTION

Maritime transport is economically central but faces safety risks from growing traffic, human error, and costly manual surveillance. The paper surveys how AIS data sources and navigation analyses can support anomaly detection, route estimation, collision prediction, and path planning.

  • Around 90% of world trade is carried by sea, while growing trade increases demand for larger and faster ships and raises maritime safety concerns.
  • Human judgement, manual surveillance, and crew fatigue contribute to safety risks, with human errors causing 75%-96% of marine accidents.
  • AIS broadcasts vessel kinematic and static information that can support intelligent analyses for maritime safety, security, and efficiency.
  • The survey covers traffic anomaly detection, route estimation, collision prediction, and path planning using AIS data.
  • AIS data quality is challenged by volume, incompleteness, noise, and rogue or dark vessels, while free sources may not guarantee completeness or timestamp resolution.
  • Commercial AIS providers offer historical or live data, but timestamp resolution is guaranteed in the sampled datasets only for VT E., despite full-precision latitude and longitude fields across providers.

D. Data Quality Comparison

The survey compares AIS providers through sampled data and evaluates availability, access, completeness, validity, time resolution, precision, and data type. It cautions that sample sizes and collection methods vary, limiting the certainty of some quality descriptions.

  • Data quality assessment used samples ranging from hundreds to millions of records, collected through provider requests, crawling, or website searches.
  • Commercial samples differed in scale: IHG., exactE., and VT E. supplied thousands of records, while MarineC. supplied millions.
  • Sample collection varied across providers, including automatic crawling of millions of Aishub records and decoding thousands of raw Ais E. NMEA feeds.
  • Position and timestamp quality are central evaluation dimensions, but descriptions may deviate because sample volumes differ and can be limited.
  • The comparison records provider availability, historical/live access, field completeness, heading validity, timestamp resolution, position precision, data type, and access method.
  • AIS anomaly detection models historical traffic normalcy to identify position, speed, and time anomalies, using geographical or parametrical model classes.

B. Geographical (Map-Dependent) Models

Geographical models learn local traffic normality over maps or spatial partitions and detect deviations in vessel movement. Their interpretability and online operation come with sensitivity to spatial design, data requirements, or omitted vessel attributes.

  • Normalcy Box: The normalcy box method partitions port areas into nested regions and learns acceptable minimum and maximum speeds from expert-labelled normal vessel histories.
  • Normalcy Box: Observations sufficiently outside learned normalcy boxes are flagged as unusual, while boxes can be updated or created and deleted efficiently as new data arrive.
  • Fuzzy ARTMAP: Fuzzy ARTMAP discretizes course into four directions and speed into three states, then learns trajectory transitions on a latitude-longitude grid.
  • Fuzzy ARTMAP: FAM is unsupervised and computationally inexpensive online, but requires substantial historical data and its accuracy and path resolution depend on grid size.
  • Holst Model: Holst models estimate local feature distributions in grid cells with GMM or KDE, enabling anomaly detection by thresholding estimated probabilities.
  • Holst Model: Holst models can fit local densities with limited human intervention, but incur high computational cost and require sufficient normalcy data; results vary with grid design.
  • Potential Field Method: PFM assigns decaying charges to vessel trace locations and forms a potential field whose high-intensity areas represent emergent normal traffic patterns.
  • Potential Field Method: PFM jointly detects spatial and temporal anomalies and is easy to visualize, but omits sailing direction and ship type, limiting use where separated traffic lanes matter.

C. Parametrical (Map-Independent) Models

Parametrical models detect maritime anomalies by learning probabilistic representations of vessel motion, with different methods trading modeling flexibility against computational cost and updateability.

  • TCM: TCM represents trajectory points with position, speed, and velocity vectors, clusters them using SOM, and models each cluster with a GMM.Bayesian inference on GMM probabilities produces anomaly likelihoods, including accumulated probabilities for persistent anomalies.
  • TCM: TCM thresholds let operators control type 1 and type 2 errors while detecting both obvious and persistent anomalies.
  • TCM: TCM can incorporate static and kinematic vessel information, but SOM and GMM impose high computational cost and complicate online model updates.
  • Gaussian Process: Gaussian-process anomaly detection predicts velocity from ship position and flags deviations between observations and the predictive distribution.GP predictions provide both an expected value and confidence through separate mean and variance functions.
  • Gaussian Process: GPs offer broad applicability and analyzable probabilistic models, but their high computational cost and poor scalability limit big-data and real-time use.Kernel-based approximations such as Kd-trees can reduce computational burden.

3) Bayesian Network (BN): 

Bayesian networks model vessel variables and their conditional dependencies to assess anomalous traffic, while the surrounding survey contrasts physical trajectory models and their practical limitations.

  • Bayesian Network: A Bayesian network represents vessel variables as nodes in a directed acyclic graph, with edges encoding conditional dependencies and local probability distributions.The network represents the joint distribution over its variables.
  • Bayesian Network: Variables such as ship type and speed are assigned discrete states, while network edges can come from expert knowledge or training data.
  • Bayesian Network: A vessel is flagged as anomalous when the mean joint probability over k consecutive observations falls below a threshold.
  • Bayesian Network: Bayesian networks incorporate expert knowledge and factors including weather, currents, vessel properties, and interactions, but depend on modeling assumptions and can be computationally expensive.
  • Curvilinear Model: Curvilinear models cover linear, circular, and parabolic ship motions, but solving their transition equations exactly is difficult and requires approximations.

2) Lateral Model:

Lateral, ship-dynamic, and neural-network models provide alternatives for vessel trajectory prediction, but their usefulness depends on parameter availability, input assumptions, and training choices.

  • Lateral Model: The lateral, or bicycle, model focuses on lateral motion while predicting longitudinal motion by integration from longitudinal acceleration.Steering angle is the lateral input, and modeled velocities support future-trajectory prediction through geometric relations.
  • Lateral Model: The lateral model covers constant steering, constant yaw rate, and constant heading, but critical parameter and input assumptions may not hold in practice.
  • Ship Model: Ship-dynamic models use vessel dimensions and hydrodynamic parameters to predict motion more accurately, but those parameters are less available for encountered vessels than for own vessels.This makes such models potentially more useful in simulation than real-world applications.
  • Neural Network Method: Neural-network trajectory prediction maps vessel static and kinematic features to future outputs and selects a network structure for the task.Training adjusts synaptic weights so network outputs approach desired values.
  • Neural Network Method: Trajectory-prediction studies report little performance difference across tested hidden-neuron counts, while the training algorithm significantly influences results.In the cited comparison, BP performs much better than Mercator.
  • Neural Network Method: Neural networks require no prior ship or weather assumptions, but training can converge slowly and lacks general rules for selecting activation functions and hidden-layer structure.

2) Gaussian Process Method:

Gaussian processes and Kalman filters are used to predict vessel trajectories from spatial motion patterns or sequential state estimates, with assumptions and model initialization shaping their practical performance.

  • Gaussian Process Method: Gaussian-process trajectory methods learn mappings from location to trajectory-derivative distributions and predict derivatives at new locations.
  • Gaussian Process Method: The Ornstein-Uhlenbeck process is a stationary Gaussian-process variant whose mean and variance do not change over time, an assumption described as strict for real applications.
  • Kalman Filter: Kalman filtering combines a dynamics model, control inputs, and sequential measurements through prediction and update phases to estimate a system state.Ship-state examples include latitude, longitude, speed, course, and acceleration, with position measurements.
  • Kalman Filter: Extended Kalman filtering combined with curvilinear motion models is used for ship trajectory estimation.
  • Kalman Filter: Kalman filters usually provide good short-term prediction accuracy, but performance depends critically on initial states and model assumptions; extended-filter solutions are generally not globally optimal.

4) Minor Principal Component:

Minor Component Analysis (MCA) estimates ship motion by using eigenvectors associated with the smallest eigenvalues, while hybrid route-estimation models combine physical and learning components. These methods offer simplicity, adaptive parameter selection, or richer motion prediction, but introduce limitations such as weak nonlinear modeling and computational or tuning costs.

  • Minor Component Analysis: MCA selects eigenvectors corresponding to the k smallest eigenvalues and solves a linear equation to predict the missing part of a ship-motion sequence.The prediction uses a window parameter and separates past data from the prediction component.
  • Minor Component Analysis: MCA was compared with neural networks, autoregressive models, and Wiener predictors, with results demonstrating it as a promising ship-motion prediction method.
  • Minor Component Analysis: MCA is simple to understand and implement but may model nonlinear motion poorly because component analysis is a linear transformation.
  • Hybrid Models: Hybrid route-estimation models combine physical motion descriptions with learning algorithms, including curvilinear models paired with extended Kalman filtering.The curvilinear model supplies the motion model, while ship-position measurements form the measurement model.
  • Hybrid Models: The curvilinear-model and extended-Kalman-filter combination can predict location, speed, and acceleration using only location data.It can predict several types of ship motion, but its performance depends strongly on initial-state and noise assumptions.
  • Hybrid Models: Multi-model hybrids use optimization components such as PSO or GA to tune learning algorithms and improve generalization with less human attendance.Their training may take longer because the optimization component repeatedly runs the learning component.

3) Other Methods:

Other route-estimation and collision-assessment methods include alternative predictors and ship-domain models ranging from fixed geometric shapes to speed- or context-dependent compounds. Ship-domain design remains an open problem because safety regions depend on vessel, traffic, environmental, and navigator factors.

  • Other Methods: Other route-estimation methods include ROT-based, stochastic linear-system, QRLCS, and sequential Monte Carlo approaches.
  • Collision Risk Assessment: Collision risk assessment determines whether route planning is invoked and can use ship-domain violations or risk indices based on SD, DCPA, and TCPA.
  • Ship Domains: Ship-domain definition is an open problem, with existing models categorized as simple, compound, or learnt domains.
  • Simple Ship Domain: Simple ship domains use fixed regular geometrical shapes, including ellipse and polygon models, with uniform mathematical formulas.
  • Compound Ship Domain: Compound ship domains vary their safety-region shape or size with conditions such as speed, course, or traffic density.Examples include a front half-ellipse and rear circle whose dimensions change with target-ship speed.
  • Compound Ship Domain: Dynamic safety-distance models can incorporate ship class, length, breadth, velocity, and relative velocity when defining domains or assessing collision risk.
  • Learnt Ship Domain: Learnt ship domains are motivated by the inability of manually defined domains to include all influencing factors, but require training samples supplied by experienced navigators.Questionnaire responses from captains and watch officers were used to train a neural network for domains varying with size, encounter angle, and speed.

3) Learnt Ship Domain:

Learnt ship-domain and collision-risk methods use neural networks, empirical AIS data, and CPA-based indices to represent vessel interactions and threat levels. More general risk indices include more influencing factors, but their complexity also increases.

  • Learnt Ship Domain: Neural-network ship-domain models learn mappings from vessel-interaction variables to threat levels or domain diameters.One model uses bearing, target and own-ship speeds, relative course, and distance; threat values range from 0.1 for safe to 0.9 for collision.
  • Learnt Ship Domain: Empirical ship-domain studies use large amounts of AIS data to estimate domains for different influencing factors during ship interactions.
  • Learnt Ship Domain: Learnt domains are flexible and self-adaptive and can incorporate different influencing factors, but require many training samples from experienced navigators.
  • CPA-Based Risk Indices: CPA is a key collision-risk indicator used to define risk indices, including indices based on DCPA and TCPA.
  • CPA-Based Risk Indices: R3 can be evaluated for any shape of encountered-ship domains and provides a dynamic risk index over time.
  • CPA-Based Risk Indices: Risk indices from R1 to R5 become more general by incorporating additional influencing factors, while their complexity increases.

2) Fuzzy Logic Method:

Fuzzy-logic methods represent navigation conditions and collision risk through graded membership rules. They support modular avoidance systems and guarding-ring risk assessment, but depend on expert-designed variables and rules.

  • 2) Fuzzy Logic Method:: Fuzzy logic represents variables with graded values between false and true using membership functions.This extends traditional Boolean logic, where variables are either true or false.
  • 2) Fuzzy Logic Method:: A fuzzy collision-avoidance system combines object detection, obstacle avoidance, trajectory tracking, and speed control modules.The module rules are determined from human intuition, with an H∞ autopilot producing control output from navigation and environmental inputs.
  • 2) Fuzzy Logic Method:: Virtual Force Field combines attractive force toward the target track and repulsive force away from obstacles, with contributions set by fuzzy relative-motion rules.The rules use relative distance, relative course, and relative speed between own ship and target ship.
  • 2) Fuzzy Logic Method:: A fuzzy ship-domain model sets the guarding-ring radius from ship length, speed, and sea condition through 27 fuzzy rules.The radical axis connects the guarding rings’ intersection points, and its length serves as the collision alert index.
  • 2) Fuzzy Logic Method:: Ships are considered safe when collision-point time discrepancy Δt exceeds τ; otherwise, the danger assessment indicates collision risk.τ comes from a fuzzy S-function model, while Δt is the time discrepancy between observations of the potential collision point.
  • 2) Fuzzy Logic Method:: COLREGs-based fuzzy assessment divides own ship’s domain into eight encounter-dependent sectors and fuzzifies target motion relative to collision conditions.The sectors cover overtaking, crossing, and head-on encounters.

VI. PATH PLANNING

AIS-based path planning seeks safer, lower-cost routes after collision risk is detected. Surveyed approaches include shortest-graph search and evolutionary optimization, balancing route optimality, maneuverability, constraints, and computation.

  • VI. PATH PLANNING: Path planning finds a safer route that minimizes costs such as time, distance, and course changes after a potential collision is detected.Manual planning can be laborious, time-consuming, and suboptimal.
  • B. Shortest Graph Path Method: Maze routing searches unvisited grid cells breadth-first using sea, arrival-time, ship-domain, and visited indicators to obtain a shortest route.The method avoids moving and static obstacles and has linear complexity without constructing an adjacency matrix.
  • B. Shortest Graph Path Method: A turning penalty replaces arrival time with generalized arrival time, discouraging routes with excessive direction changes in dense obstacle settings.Direction differences of 45, 90, and 135 degrees receive parameterized penalties, while same-direction gates receive zero penalty.
  • B. Shortest Graph Path Method: A* methods incorporate vessel physical constraints and COLREGs, trading faster large-graph search for potentially sub-optimal paths.Breadth-first maze, Lees, or Dijkstra searches are described as suitable for smaller or medium graphs and capable of global optima.
  • B. Shortest Graph Path Method: Shortest-path methods are simple and can guarantee an optimal path when one exists, but require rectangular grids and accurate obstacle and destination information.Those exact inputs are often difficult to obtain.
  • Evolutionary Algorithm: Evolutionary algorithms encode paths as chromosomes, evaluate fitness and constraints, select survivors, and apply crossover, mutation, and variation.Coordinate pairs can represent path genes, while objectives may combine danger, distance, straightness, energy, and navigation rules.
  • Evolutionary Algorithm: Evolutionary objectives can combine safety, distance, smoothness, and sailing time through weighted cost terms, while incorporating static and dynamic vessel information.Other formulations optimize collision-avoidance restoration cost under turning-angle and restoration-time constraints.
  • Evolutionary Algorithm: Evolutionary algorithms accommodate diverse obstacles and multi-target planning, but their high computational cost and slow convergence hinder real-time use.The survey also describes variants incorporating COLREGs, weather effects, motion-model parameters, and Bayesian trajectory probabilities.

D. Evolutionary Set Method

Evolutionary Set methods operate directly on ship trajectories rather than encoded chromosomes. This reduces computational overhead and supports simultaneous optimization of trajectories for multiple encountered ships.

  • D. Evolutionary Set Method: Evolutionary Set methods apply evolutionary operations directly to sailing paths, eliminating the encoding and decoding process used in traditional evolutionary algorithms.An individual may contain a set of trajectories for ships involved in an encounter.
  • D. Evolutionary Set Method: Evolutionary Set operators include reproduction, mutation, specialized trajectory improvement, and validation or fixing adjustments.Mutation can insert, join, shift, or delete nodes, while validation evaluates trajectories against moving ships and static obstacles.
  • D. Evolutionary Set Method: ESoSST extends the method with modified optimization criteria and COLREGs-violation penalties, while later work focuses on COLREGs Rule 10 and Traffic Separation Schemes.These extensions modify or replace evolutionary mechanisms to improve performance.
  • D. Evolutionary Set Method: Direct trajectory operations give Evolutionary Set methods much less computational cost and reported faster execution than traditional evolutionary algorithms.They can search optimal trajectories for all encountered ships simultaneously rather than only for own ship.

E. Artificial Potential Field

Artificial potential field methods guide ships toward destinations while repelling them from obstacles, offering dynamic routing for collision avoidance. Their realtime flexibility is offset by frequent course adjustments and difficulties modelling feasible ship forces and repulsion.

  • E. Artificial Potential Field: Artificial potential fields attract ships toward target destinations and repel them from obstacles through a combined virtual force.The ship follows the gradient direction of the total potential field, with attraction and repulsion represented separately.
  • E. Artificial Potential Field: Improved variants incorporate ship velocity and maneuvering behavior, while harmonic potential fields address the local-minima limitation.The harmonic formulation uses boundary assumptions and satisfies the Laplacian equation in the potential domain.
  • E. Artificial Potential Field: The method provides a dynamic route at every moment, making realtime capability its principal advantage.Route calculations are updated continuously as the ship moves through the environment.
  • E. Artificial Potential Field: Frequent course adjustments can be time-consuming and dangerous for ships, while exact driven-force production and realistic repulsive-force modelling remain difficult.The resulting total force is sensitive to how repulsive forces are specified.
  • Survey context: AIS navigation research spans anomaly detection, route estimation, collision prediction, and path planning, with surveyed methods retaining important trade-offs and open challenges.The survey contrasts model intuitiveness, information use, prediction complexity, ship-domain design, and performance under idealized conditions.
Loading 1606.00981v1…