Source-linked AI summary
Online Event Recognition from Moving Vessel Trajectories
Kostas Patroumpas, Elias Alevizos, Alexander Artikis, Marios Vodas, Nikos Pelekis, Yannis Theodoridis
TL;DR
Maritime surveillance needs to process large, continuous AIS streams, yet existing work does not provide a streaming framework tailored to noisy and intermittent geostreaming from large fleets. This paper combines online trajectory summarization with complex-event recognition, retaining critical points at close to 98% compression with negligible approximation-quality loss and supporting alerts to marine authorities.
Problem
Maritime surveillance requires scalable, continuous spatiotemporal processing of AIS streams, and the authors identify no streaming framework specifically tailored to fluctuating, noisy, intermittent messages from large fleets.
Source: §1 · pp. 2–3Method
Results
The online trajectory summary achieves close to 98% data compression with negligible loss in approximation quality.
Source: §1 · p. 2Takeaways & Limitations
Critical-point summaries support recognition of dangerous or suspicious situations, with alerts issued to marine authorities.
Source: §1 · p. 2The tracker assigns only one crisp annotation to each critical point; the authors plan to explore fuzzy, probabilistic multiple annotations.
Source: §3.2 · p. 10
Abstract
from arXiv · showhide
We present a system for online monitoring of maritime activity over streaming positions from numerous vessels sailing at sea. It employs an online tracking module for detecting important changes in the evolving trajectory of each vessel across time, and thus can incrementally retain concise, yet reliable summaries of its recent movement. In addition, thanks to its complex event recognition module, this system can also offer instant notification to marine authorities regarding emergency situations, such as risk of collisions, suspicious moves in protected zones, or package picking at open sea. Not only did our extensive tests validate the performance, efficiency, and robustness of the system against scalable volumes of real-world and synthetically enlarged datasets, but its deployment against online feeds from vessels has also confirmed its capabilities for effective, real-time maritime surveillance.
1 Introduction
Maritime surveillance must process continuous AIS streams at large scale, while existing streaming systems and vessel-interaction approaches do not address the full real-time maritime setting. The proposed system combines online trajectory summarization with complex-event recognition, and evaluates its performance, scalability, and real-world deployment.
1 Introduction
- AIS data from more than 580,000 vessels worldwide creates demand for scalable, continuous spatiotemporal processing over massive streams.AIS was initially intended to help prevent collisions, while its volume, precision, and real-time availability also support broader monitoring applications.Source: §1 · p. 2
- The system tracks major trajectory changes, retains critical points with close to 98% compression and negligible approximation-quality loss, and uses them to detect dangerous or suspicious maritime events.Detected situations include fast-approaching vessels and package picking at open sea; the system can issue alerts to marine authorities.Source: §1 · p. 2
- The authors identify a gap in streaming frameworks tailored to fluctuating, noisy, intermittent AIS geostreams from large fleets, beyond work focused on predictive accuracy or generic stream processing.The related vessel-interaction approach cited focuses more on predictive accuracy than real-time streaming performance.Source: §1 · pp. 2–3
- The paper adds single-pass AIS noise-reduction heuristics, online trajectory filters, and rules for efficiently recognizing maritime events involving vessels and geographical zones.It is an extended and revised version of earlier work, with improved algorithms and more thorough empirical validation.Source: §1 · p. 3
- The evaluation examines real vessel traces, scalable synthetically enlarged datasets, and deployment against online AIS feeds collected across the Aegean Sea.The paper reports investigation of system robustness, efficiency, and timeliness, along with event recognition related to sensitive environmental zones.Source: §1 · p. 3
2 System Architecture
The architecture processes streaming AIS positions through noise reduction and online trajectory tracking, then forwards annotated critical points to complex-event recognition. Its output supports real-time alerts about potentially dangerous or suspicious maritime situations.
2 System Architecture
- The system consumes AIS tracking messages, detects vessel movement features, and recognizes complex events such as suspicious activity.The processing flow is depicted in Figure 1.Source: §2 · p. 3, Figure 1 · p. 4
- Incoming positions are filtered for noise before the Mobility Tracker checks how vessel velocity changes over time.The system accounts for noise from sea drift, delayed messages, and GPS discrepancies; the tracker operates on clean data.Source: §2 · p. 4
- At each window slide, detected trajectory events are compiled into annotated critical points, far fewer than the original positions, for real-time characterization and visualization.The architecture uses sliding windows to keep pace with newly arriving stream tuples; successive windows may overlap when the slide step is shorter than the window range.Source: §2 · pp. 3–5
- The trajectory synopsis annotates events such as turns and stops, adding semantic information to simplified traces rather than only reducing positional updates.The paper contrasts this feature with dead-reckoning policies, which may be unsuitable because maritime control centers wish to locate ships frequently.Source: §2 · p. 5
- Complex-event recognition combines critical points with static geographic data, such as protected areas, and pushes recognized suspicious or dangerous situations to marine authorities in real time.Examples include loitering and vessel pursuit.Source: §2 · p. 5
3 Detecting Trajectory Events
The system filters noisy AIS positions, detects movement changes in each vessel’s stream, and retains critical points as a concise trajectory synopsis. Its tracker distinguishes instantaneous events from patterns that accumulate over time.
3 Detecting Trajectory Events
- Movement events such as stops, turns, and slow motion mark critical points that provide a concise, reliable representation of each vessel’s course.Source: §3 · p. 5
- Instantaneous events describe individual positions, while long-lasting events combine a bounded sequence of events to detect cumulative changes in motion.Source: §3 · p. 6
3.1 Online Noise Reduction
- Single-pass noise filters use instantaneous velocity and the vessel’s recent mean velocity to flag out-of-course positions, abrupt turns, implausible acceleration, and duplicate or contradictory reports.Source: §3.1 · pp. 7–8
- 20% of raw AIS positions may qualify as noise; noise reduction benefits performance without sacrificing accuracy in the resulting approximation.Source: §3.1 · p. 8
3.2 Online Tracking of Moving Vessels
- The tracker identifies reporting gaps, long-term stops, slow motion, and smooth turns from buffered positions, then emits annotated critical points for detected events.Source: §3.2 · pp. 9–10
- Each critical point receives only one annotation, so cases that could qualify as both a gap and a stop cannot retain both labels.Source: §3.2 · p. 10
- Instantaneous-event and communication-gap checks are O(1) per incoming position; long-lasting-event checks are O(b), and the tracker is reported to operate near real time.Source: §3.2 · p. 11
- The compressor evicts noncritical locations, while parameter choices trade trajectory-reduction efficiency against approximation accuracy.Source: §3.2 · p. 11
4 Complex Event Recognition
The complex-event recognition module combines movement events with port and protected-area locations to identify potentially suspicious or dangerous maritime situations. It uses RTEC to specify and recognize temporal patterns with spatial constraints.
4 Complex Event Recognition
- The module combines movement events with port and protected-area locations to recognize potentially suspicious or dangerous maritime situations.The trajectory detector supplies critical movement events to the complex-event recognition module.Source: §4 · p. 12
4.1 Event Calculus for Run-Time reasoning
- RTEC represents fluent values over maximal time intervals, deriving each interval from an initiation and the first subsequent termination.Its rules describe event instances, event effects, and fluent values; the gap example initiates a communication-gap fluent when a vessel stops transmitting away from ports.Source: §4.1 · p. 13
4.2 Spatial Indexing
- A grid assigns ports and areas of interest to overlapping cells, narrowing proximity and point-in-polygon checks to vessels and areas in the same or adjacent cells.Areas and ports are assigned to cells offline; a vessel’s cell is determined before each recognition query.Source: §4.2 · p. 14
4.3 Specifying Complex Maritime Events
- A suspicious delay is an open-sea communication gap whose estimated maximum vessel speed falls below a threshold, assuming straight-line travel between gap endpoints.The paper notes that a more refined estimate would account for routes that are impossible or unlikely under terrain or weather conditions.Source: §4.3 · p. 15
- RTEC can express hierarchical event patterns: simultaneous suspicious delays by two vessels in the same area may indicate a rendezvous, but not its precise place or time.The system lacks vessel locations during communication gaps.Source: §4.3 · pp. 15–16
- A fast approach is recognized when a vessel changes speed at open sea, exceeds 20 knots, and heads toward at least one nearby vessel.Domain experts selected the 20-knot threshold.Source: §4.3 · p. 16
- The package-picking pattern matches a vessel that stops at open sea and later another vessel stopping within half a kilometer of that location.The rule describes the second vessel arriving within no more than an hour after the first vessel resumes moving.Source: §4.3 · pp. 16–17
4.4 Recognizing Complex Maritime Events
- RTEC recognizes events from movement-event occurrences within a sliding window, retaining late-arriving events when the window range exceeds the query slide step.Events before or at the window’s lower boundary are discarded; delayed events in the still-open window can be considered at a later query.Source: §4.4 · p. 17
5 Empirical Evaluation
The evaluation uses a three-month AIS dataset of vessels in the Aegean, Ionian, and part of the Mediterranean, replayed as a stream and supplemented with enlarged data for stress tests.
5 Empirical Evaluation
- The experiments evaluate the trajectory tracker and RTEC, the complex-event recognition component, using the real and synthetically enlarged data streams.Simulation settings are listed in Table 2; mobility-tracking parameters are listed in Table 3.Source: §5 · pp. 17–18, Table 2 · p. 18, Table 3 · p. 18
- The real AIS dataset contains 23 GB of messages from 6,425 vessels over three months and yields 168,240,595 cleaned, timestamped positions.The messages span 1 June to 31 August 2009; many vessels were active only briefly, while others sailed frequently.Source: §5 · pp. 17–18
- Replay follows the original timestamps and fluctuating vessel activity, averaging about 50 positions per second during active periods.Additional experiments increased the arrival rate to stress-test the system, including complex-event streams with 1.2 million vessels and 3.2 billion movement events.Source: §5 · p. 18
5.1 Assessment of Trajectory Detection
The trajectory tracker processes streaming AIS positions into critical movement points, producing compact trajectory summaries while measuring processing cost, approximation error, and compression quality.
5.1.1 Performance of online mobility tracking
- For a 24-hour window and 4-hour slide, tracking takes less than 2 seconds per batch while validating almost 30,000 fresh positions.For windows up to 2 hours, the worst-case tracking time is under 150 ms per incoming batch; costs increase with larger slide steps and batches.Source: §5.1.1 · p. 19
5.1.2 Performance under varying arrival rates
- At 10,000 positions per second, the tracker processes 600,000 fresh positions per minute and reports results in less than 13 seconds, before the next window slide.The experiment used a 10-minute window and a 1-minute slide.Source: §5.1.2 · pp. 20–21
5.1.3 Approximation error
- At a 20° turn threshold, average RMSE is 22 meters and maximum RMSE is 133 meters; both error measures rise as the threshold is relaxed.Across vessel trajectories, only three vessels have RMSE above 100 meters; thresholds of 10° or 15° may balance compression and retained mobility detail.Source: §5.1.3 · p. 22, Figure 11 · p. 22
5.1.4 Compression efficiency
- Compression remains above 92% and reaches 98% with a relaxed turn threshold, leaving only 2% of original locations as critical points.The relaxed threshold mostly removes local manoeuvres with little impact on the vessel’s course; filtering noise also reduces retained points.Source: §5.1.4 · p. 23, Figure 12 · p. 23
- With increasing communication-gap threshold, gap points are reduced while compression remains at least 96%.Even at a five-minute threshold, additional points are needed to mark lost contact.Source: §5.1.4 · p. 23, Figure 13 · p. 23
5.1.5 Quality of synopses
- Normal positions make up more than half of raw locations, while turns comprise 60% of critical points despite representing only 1.5% of raw locations.Communication gaps account for almost 32% of critical points, while long-term stops are 4.5% after successive pause events are merged at their centroid.Source: §5.1.5 · pp. 23–24, Figure 15 · p. 24, Figure 14 · p. 24
5.2 Assessment of Complex Event Recognition
RTEC recognizes maritime complex events from movement events and geographic data, with experiments examining grid choice, parallel processing, irrelevant input, and enlarged streams.
5.2 Assessment of Complex Event Recognition
- RTEC combines movement-event predicates with protected-area polygons and port locations to recognize six maritime event types.The recognized types include illegal shipping, suspicious vessel delay, rendezvous, suspicious areas, vessel pursuit, and package picking.Source: §5.2 · p. 25
5.2.1 Grid partitioning
- The 30×30 grid was selected because it gives a stable number of recognized events with time performance comparable to the 10×10 grid.The 90×90 grid is almost twice as slow as the 10×10 grid, while all tested grid configurations remain under three seconds on average.Source: §5.2.1 · pp. 25–26, Figure 16 · p. 26
5.2.2 Performance under varying window sizes and distributed configurations
- With a six-hour window, parallel recognition takes 0.4 seconds on twelve processors, compared with 14 seconds on one processor.The evaluation includes windows from one to 24 hours and two-, four-, and twelve-processor configurations, using compact rectangular sub-grids.Source: §5.2.2 · pp. 26–27, Figure 17 · p. 27
5.2.3 Tolerance to irrelevant events
- Recognition time and memory use remain robust as irrelevant normal events increase, while the average number of recognized complex events stays the same.The test stream ranges from 16 million critical movement events alone to those events plus 262 million normal events.Source: §5.2.3 · pp. 27–28, Figure 18 · p. 28
5.2.4 Artificially enlarged data streams
- With twelve processors, RTEC handles streams about 70 times larger for a 60-minute window and 160 times larger for a 10-minute window while meeting a 60-second response limit.For the 60-minute window, the reported setting covers approximately 450,000 vessels, about 400,000 movement events, around 105,000 recognized complex events, and 4.9 GB of memory.Source: Figure 20 · p. 30, §5.2.4 · p. 31
6 System Deployment
The system was deployed on live AIS messages collected across the Aegean, where it processes vessel positions and archives or exports detected critical points. The deployment has been active since April 2015.
6 System Deployment
- The Aegean deployment has operated on real-time AIS messages since April 2015, confirming integration with an online data source for locating and correlating vessel-trace events.Fresh records are typically fetched hourly, up to 40,000 per hour; exported points currently lag by six hours because they are exported after leaving the sliding window.Source: §6 · p. 32
7 Summary & Future Work
The system compresses vessel trajectories while supporting real-time complex-event recognition at large scale. Future work targets adaptive parameterization, event-pattern construction and uncertainty, additional data sources, and transfer to other mobility streams.
7 Summary & Future Work
- Almost 98% compression is reported with tolerable approximation error, while CE recognition runs in real time on datasets reaching 1,28M vessels and 3,2B critical positions.Trajectory detection is also described as highly scalable and readily parallelizable.Source: §7 · p. 33
- Planned trajectory-detector work targets parameter sensitivity through adaptive, auto-calibrated settings keyed to vessel size, type, and motion patterns.Source: §7 · p. 33
- Planned event-pattern work would automate construction through incremental structure learning and address corrupt AIS messages and incomplete patterns with a Markov Logic Networks dialect.Source: §7 · p. 33
- The authors plan to add heterogeneous data sources and test whether the methodology transfers to aircraft or vehicle geostreaming, despite differences in event definitions and configurations.Additional sources may improve monitoring quality and support recognition of additional events.Source: §7 · p. 33