SLIDE 1 How not to drown in a sea of information: An event recognition approach
Elias Alevizos1, Alexander Artikis2,1, Kostas Patroumpas2,3, Marios Vodas2, Yannis Theodoridis2, Nikos Pelekis2
1NCSR Demokritos 2University of Piraeus 3National Technical University of Athens
November 1, 2015
SLIDE 2
Goals & Motivation
◮ 580,000 AIS-equipped vessels worldwide ◮ Distributed sources ◮ Ambiguous/uncertain messages ◮ Build a system monitoring the activity of thousands of vessels,
transmitting AIS messages every few seconds.
◮ Incorporate spatial, background knowledge (thousands of
special areas of interest).
◮ Real-time performance. ◮ Scalability with enlarged datasets.
SLIDE 3
Motivating example (1): Fast Approach
◮ A vessel is moving at a high speed ... ◮ towards another vessel ... ◮ away from ports.
SLIDE 4
Motivating example (2): Package Picking
◮ A vessel stops in a location ... ◮ and another vessel stops in the same location ... ◮ in a short period of time.
SLIDE 5
Motivating example (3): Possible Rendezvous
◮ Two vessels are suspiciously delayed ... ◮ in the same location ... ◮ at the same time.
SLIDE 6
Complex Event Recognition
SLIDE 7 Complex Event Recognition for Maritime Surveillance
Input:
◮ Raw input: 168M real-world AIS messages (June–August
2009)
◮ Input to event recognition engine
◮ ≈ 16M critical movement events from 6,5K vessels (95%
compression w.r.t. to raw data).
◮ Spatial knowledge ◮ Surveillance area (720 km x 900 km), divided into grid cells. ◮ > 4K areas of interest (e.g. NATURA), with ≈ 78K edges. ◮ 64 ports.
Output:
◮ Spatio-temporal events/activities of interest (e.g. package
picking).
SLIDE 8 Approach: Run-Time Event Calculus
◮ A logic programming language for representing and reasoning
about events and their effects
◮ formal, declarative semantics.
◮ Succinct representation of complex temporal phenomena
◮ Intuitive representation → facilitates interaction with domain
experts unfamiliar with programming.
◮ Support for intervals. ◮ Support for hierarchical knowledge.
◮ Combination of streaming data with background atemporal
knowledge.
SLIDE 9
Empirical Evaluation
(a) Average recognition
times.
(b) Average number of
Movement Events (input).
(c) Average number of
Complex Events (output).
◮ ≈ 16M Movement Events, 6,5K vessels, 4K areas (78K
edges), 900 cells (720 km x 900 km)
◮ Real-time performance
SLIDE 10
Summary
◮ Sustain large amounts of streaming messages from vessels. ◮ ... from real data. ◮ Real-time response. ◮ Patterns motivated from real-world demands (to work on
refinement).
SLIDE 11 Current and future work
◮ Performance
◮ Larger datasets ◮ Test different partition schemes ◮ More efficient parallelization (rewrite engine in Spark)
◮ Accuracy
◮ Pattern refinement ◮ Include historical, statistical, contextual data ◮ Probabilistic extensions ◮ Ground truth?
SLIDE 12
Current and future work
(d) Average recognition
times.
(e) Average number of
Movement Events (input).
(f) Average number of
Complex Events (output).
Experiments with synthetically enlarged datasets
◮ Increase factor 20–200, 320M–3,2B Movement Events,
128K–1,28M vessels, 4K areas.
