Increasing Maritime Situation Awareness via Trajectory Detection, Enrichment and Recognition of Events

George A. Vouros, Akrivi Vlachou, Georgios M. Santipantakis, Christos Doulkeridis, Nikos Pelekis, Harris V. Georgiou, Yannis Theodoridis, Kostas Patroumpas, Elias Alevizos, Alexander Artikis, Georg Fuchs, Michael Mock, Gennady L. Andrienko, Natalia V. Andrienko, Christophe Claramunt, Cyril Ray, Elena Camossi, Anne-Laure JousselmeW2GIS 2018 · DOI

TL;DR

This paper presents the datAcron project's approach to enhancing maritime situation awareness by processing vast amounts of real-time and historical vessel data. It introduces a big data architecture that integrates, enriches, and analyzes trajectory data to detect and forecast complex events at sea, aiming to improve safety and efficiency in maritime operations.

What Problem Does This Paper Address?

The maritime sector is experiencing a significant increase in traffic and resource exploitation, leading to a surge in data generated by various sources like Automatic Identification Systems (AIS). Traditional systems struggle to handle this volume, variety, and velocity of data, hindering effective Maritime Situation Awareness (MSA). MSA requires correlating diverse and often uncertain data to understand the current and predict future maritime scenarios. This paper tackles the challenge of building a system capable of processing this "big data" in real-time to detect and forecast critical events, such as potential collisions or illegal fishing activities, thereby improving safety and supporting sustainable maritime development.

Methods

The core of the solution is the datAcron system architecture, which combines elements of the Lambda architecture but with distinct purposes for its real-time and batch layers.

  • Real-time Layer: This layer focuses on processing streaming surveillance data. Key components include:
  • In-situ Processing: Cleans and enriches incoming data with derived attributes (e.g., speed, acceleration) and detects low-level events like entering/leaving areas. This is done "in-situ" to minimize latency.
  • Synopses Generation: Compresses raw trajectory data by identifying and retaining only "critical points" that signify changes in motion (stops, slow motion, heading/speed changes, communication gaps). This dramatically reduces data volume while preserving essential information.
  • RDF Generation and Data Integration: Converts trajectory synopses into RDF (Resource Description Framework) format, enabling integration with other data sources (e.g., weather, regulations) to create a knowledge graph.
  • Link Discovery: Identifies spatio-temporal and proximity relationships between entities (vessels, areas) within the RDF graph, enriching the data with contextual information.
  • Complex Event Recognition and Forecasting: Uses Deterministic Finite Automata (DFA) to define event patterns and converts them into Pattern Markov Chains (PMC) for probabilistic forecasting of complex events.
  • Batch Layer: This layer handles persistent storage of enriched trajectories and other data for offline analysis, including trajectory clustering and visual analytics.

The system leverages technologies like Apache Flink for stream processing and Apache Spark for batch processing, with Apache Kafka for inter-component communication.

Datasets

The paper mentions using real vessel data obtained through the Automatic Identification System (AIS). Specifically, it refers to a dataset containing approximately 18 million AIS positions transmitted by about 5,000 vessels sailing in the Atlantic Ocean around the port of Brest, France, between October 2015 and March 2016. Navigation features like areas of interest, coastlines, and port locations were also considered. No new benchmarks were introduced; rather, existing data was used for evaluation.

Key Insights and Ideas

  • Trajectory Synopses for Efficiency: The idea of compressing trajectories into "synopses" by only retaining critical points is a clever way to handle the massive data volume without losing crucial motion information. This significantly reduces the computational burden for real-time analysis.
  • RDF for Data Integration: Using RDF as a common data model allows for flexible integration of diverse maritime data sources, creating a rich knowledge graph that can be queried for complex relationships.
  • Spatio-temporal Link Discovery for Context: Moving beyond simple proximity, the link discovery component's ability to identify spatio-temporal relationships is crucial for understanding how vessels interact with their environment and each other, which is key for event detection.
  • Probabilistic Forecasting with Markov Chains: The conversion of event patterns into Markov Chains for forecasting is a robust approach to predict future events with associated probabilities, enabling proactive decision-making.
  • Learning from Examples: The system's ability to learn patterns leading to complex events (e.g., a sequence of "close" events leading to a "within" event) demonstrates a powerful capability for adapting to and understanding nuanced maritime behaviors.

Key Results

The paper presents preliminary results and demonstrates the functionality of the datAcron architecture.

  • Event Detection and Forecasting: The system successfully detected and forecasted events, illustrated by an example of predicting when a vessel is expected to enter a specific area.
  • Visualization: A visualization interface was developed to display low-level events, critical points, and complex events, showing ship tracks and real/predicted events. Figure 3 shows an example of event forecasting visualization, indicating predicted entry times and probabilities for vessels approaching an area.
  • Data Compression: The synopses generation module achieved significant data reduction, with compression ratios reaching up to 99% in some cases without harming the quality of the derived synopses.
  • Accuracy in Forecasting: For the "within" event use case, results showed that as a vessel gets closer to an area, the precision of the forecast increases, and the forecast intervals become smaller. The system also learned to refrain from producing forecasts for routes that do not typically lead to the event.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that the presented results are preliminary.
  • Challenges remain in discovering interactions among moving ships in a timely manner.
  • Efficient querying of very large knowledge graphs for online and offline analytics is still an area of development.
  • Cross-streaming synopses generation at the data integration level is a complex task.
  • Future Work:
    • Long-term online full trajectory predictions.
  • Improvements in forecasting complex events, including learning and refining patterns by exploiting examples.
  • Further development of the visualization interface to support human interaction.

Summary updated

BibTeX
@inproceedings{vouros2018increasing,
  author = {George A. Vouros and Akrivi Vlachou and Georgios M. Santipantakis and Christos Doulkeridis and Nikos Pelekis and Harris V. Georgiou and Yannis Theodoridis and Kostas Patroumpas and Elias Alevizos and Alexander Artikis and Georg Fuchs and Michael Mock and Gennady L. Andrienko and Natalia V. Andrienko and Christophe Claramunt and Cyril Ray and Elena Camossi and Anne-Laure Jousselme},
  title = {{Increasing Maritime Situation Awareness via Trajectory Detection, Enrichment and Recognition of Events}},
  booktitle = {W2GIS},
  year = {2018},
  doi = {10.1007/978-3-319-90053-7\_13},
}
Figure