Copyright (C) 2009 DENSO IT LABORATORY, INC. All Rights Reserved.
An event detection method using floating car data
Osamu Masutani, Hirotoshi IwasakiDenso IT Laboratory, Inc.
Copyright (C) 2009 DENSO IT LABORATORY, INC. All Rights Reserved. 2/17
SummaryBackgroundConceptMethodology- Factorization- Detection- Prediction
Conclusions
Copyright (C) 2009 DENSO IT LABORATORY, INC. All Rights Reserved. 3/17
Ex. Festivals, sports eventAs a destination- Attractive POI- Drivers would like to know “fresh”
information about events. As an obstacle- Repellent POI.- Drivers would like to know risk of
congestion caused by an event. Event information- Place and time
Background: Event
!
Event
Our aspect
Copyright (C) 2009 DENSO IT LABORATORY, INC. All Rights Reserved. 4/17
Background: Avoidance of event congestionTraffic information service- Watch and avoid→
Congestion-aware route guidance
→
Traffic prediction
Problem- Most of prediction method based on
stationery situation of traffic.- Event is irregular and unpredictable.
Event “plan” information can improve prediction accuracy.- Event-aware traffic prediction
Event ahead
18:00
Google maps
Copyright (C) 2009 DENSO IT LABORATORY, INC. All Rights Reserved. 5/17
Background: Event databaseManually collected database- For web, telematics services
Not integrated- Most of services are dedicated to
certain genre (ex. business event, leisure event)
- Private event information isn’t provided (ex. school event)
Not real-time- Re-schedule isn’t always tracked
Fireworks in Stockholm
Copyright (C) 2009 DENSO IT LABORATORY, INC. All Rights Reserved. 6/17
Concept: Event detection method with FCDFully automatic detection- Extracted from floating car data (FCD)
Integrated- Can detect any type of event.- Private event also be detected.
Real-time- Based on real-time fluctuation of FCD
Copyright (C) 2009 DENSO IT LABORATORY, INC. All Rights Reserved. 7/17
Concept: Relationship between FCD and an event
Definition- Zone : gridded areas- Attraction : Incoming cars to the zone- Generation : Outgoing cars from the
zone
Event specific pattern: Attractionincrease before event and generation increase after event
GenerationAttraction
Generation
Attraction
Stadium
Event period
Assumption
Count of attraction and generation of cars to/from an event venue indicates event-specific pattern.
Copyright (C) 2009 DENSO IT LABORATORY, INC. All Rights Reserved. 8/17
Concept: 3D view of attraction and generationSpatio-temporal density view reveals the event-specific pattern- Attraction peak (blue chunk) followed
by generation peak (red chunk)
Density mapping by iso-surface- More interpretable than point scatter
Stadium
Similar method with “crime mapping”
Nakaya, T., Yano, K., (2008). Spatio-temporal three-dimensional mapping of crime events:visualizing spatio-temporal clusters of snatch-and-run offences
Events
Copyright (C) 2009 DENSO IT LABORATORY, INC. All Rights Reserved. 9/17
Factorization: For accurate detection
An event venue isn’t identical with a grid zone- Some sets of zones are related
with one event
Event specific pattern might be hidden in stationery traffic- Various factors of fluctuation are
coincident in a zone- Event related factor should be
properly separated
Stadium
Stationary fluctuation
Event fluctuation
Copyright (C) 2009 DENSO IT LABORATORY, INC. All Rights Reserved. 10/17
Factorization: Spatio-temporal factorizationPCA on spatio-temporal space- Factor is a pair of zone and time
series.Factored zones are zones which have coincident traffic.Factored time series are separated according to its pattern.
Similar method with image factorizationOliver, N. M., Rosario, B., Pentland, A. P., (2000).A Bayesian Computer Vision System for Modeling Human Interactions
Copyright (C) 2009 DENSO IT LABORATORY, INC. All Rights Reserved. 11/17
Factorization: Factors for density of attractionCategorization of factors- Single zone / set of zones- Periodic / irregular /
transitory fluctuation
Factor 1 : Weekday morning peaks on central station and business area
Factor 2 : Evening peaks on night spotFactor 3 : Transitory peaks on a sports centerFactor 4 : Weekday midnight peaks on central
stationFactor 5 : Transitory peak on a spot
Copyright (C) 2009 DENSO IT LABORATORY, INC. All Rights Reserved. 12/17
Factorization: Event related factorZones: nearby area of the stadiumTime series: corresponding to the dates of baseball games
Weighted combination of zones
Stadium
: baseball games
Copyright (C) 2009 DENSO IT LABORATORY, INC. All Rights Reserved. 13/17
Detection: Event pattern detection by HMM
Event “state” definition- 4 states (non, pre, in, post)- Pre- and post- event states defined as
2 hours before and after the event period.
Detection of pre-event can be regarded as prediction of event4 state HMM- Trained by pair of actual event data and
FCD
In-event
Pre-event Post-event
Non-event
Copyright (C) 2009 DENSO IT LABORATORY, INC. All Rights Reserved. 14/17
Detection: Detection results
Evaluation on the stadium events- 1.5 month, near a stadium, 17
events (baseball games)
Results confirms our method works well- Event states detection:
precision = 86%- Event occurrence detection:
precision = 100%Recall = 100%
Event states and raw data
Detected events
Copyright (C) 2009 DENSO IT LABORATORY, INC. All Rights Reserved. 15/17
Normal
Prediction: Event aware traffic predictionBenefit of event information- Based on statistical prediction- Enhanced by adding a prediction pattern for event
Holiday Holiday
Event
WeekdayWeekday
Event aware
Copyright (C) 2009 DENSO IT LABORATORY, INC. All Rights Reserved. 16/17
Prediction: Evaluation resultEvaluation - Comparison event aware and normal (baseline) prediction- 1.5 month, near a stadium, 17 events (baseball games)
Event information improves prediction accuracy- Overall accuracy of event-aware prediction outperforms 38% over
normal prediction
Event period
Stadium and traffic
38%
Copyright (C) 2009 DENSO IT LABORATORY, INC. All Rights Reserved. 17/17
ConclusionsFCD attraction and generation statistics is good clue to event occurrenceFactorization and HMM can detect relatively large scale eventsEvent can refine traffic prediction abilityFuture works- Smaller events- Event classification