Traffic Incident Detection Using Probabilistic Topic Model

Traffic Incident DetectionUsing

Probabilistic Topic Model

Akira Kinoshita1, Atsuhiro Takasu2, and Jun Adachi2

<[email protected]>

1The University of Tokyo2National Institute of Informatics, Japan

International Workshop on Mining Urban DataAthens, Greece, March 28, 2014

2

Abstract

• Detect traffic incidents by comparing current traffic with usual traffic with less knowledge of traffic engineering.

• Show the method works well on the Shuto Expressway in Tokyo, using real probe-car data

• Possible by-product: characteristic analysis of road

2014-03-28A. Kinoshita, A. Takasu, and J. Adachi.Traffic Incident Detection Using Probabilistic Topic Model

3

Background


東池袋Higashi-

Ikebukuro

護国寺Gokokuji

早稲田Waseda

Local slowdown Incident？

Causes of congestionin Japan [E-NEXCO]

http://www.e-nexco.co.jp/english/business_activities/expressway_management/eliminating.html

Probe-car data(position+timestamp)

Congestion≠Incident

4

Related Work


[Zhu, et al. 2009]

Time

Link

slowdown@inciden

t

Date

slow slow fast

Feature vectorbased on the speed change

Distance-basedoutlier detectionafter filterling

+

−

5Research Problem& Solution Strategy

Problem• Congestion is not always caused by traffic incidents

speed reduction often occurs without any incidents

• Real-time data stream processing

Solution strategy• Regard incident as “sudden event different from usual”

• Estimate traffic state using probe-car data (or else)

• Compare current traffic with usual traffic current traffic should be much different from usual


6

Outline

Background• Automatic incident detection• Related work & research problem

Methodology• Traffic state model based on topic model• Estimate usual/current traffic states and then compare

them

ExperimentUsing probe-car data on three routes of the Shuto Expressway in Tokyo during 2011

Discussion• Results of the experiment• Possible analysis using traffic state model


7Traffic State Model (TSM) Similar to LDA [Blei 2003]

Distribution of is mixture distribution


Inbound expressway,midnight

Inbound artery,weekday morning

TrafficStates

Segment

ObservedValue

⇔Topic

⇔Word

⇔document

・・・

Congested StopModerateGood

8

Parameter Estimation

Most-likelihood estimation


9

Incident Detection Method


G M C S G C M C S G

G M C S G G G G G GUsualTraffic State

CurrentTraffic State

ObservedValue

Segment

G: GoodC: Congested

M: ModerateS: Stop

Incident

Divergence sum of divs

Toll gate

Divergence betweenusual - current states

measure by KL divergence

10

Outline



them




11

Experiment

PurposeTo show the proposed method works more efficiently than an existing method.

Data Sources• Probe-car data in the Shuto Expressway (Shutoko) in

Tokyo during 2011.Timestamp, position (longitude, latitude),vehicular speed (non-negative integer)

• Traffic log by the road administrator as the ground truth of incidents.

“Incident” includes the five events:accident, broken-down car, fallen object,

construction, looking-aside driving


12

Dataset & Preprocessing


Preprocessing1. Map matching2. Trajectory identification3. Interpolation4. Labeling

Parameter assumptions• K = 8• N = 10• Poisson distribution

Area Map[Shuto Expwy.]

(5)Ikebukuro

(4)Shinjuku

(3)Shibuya

CentralTokyo

Shuto Expwy.Routes

(3)Shibuya (4)Shinjuku (5)Ikebukuro

Inbound Out- In- Out- In- Out-

Period 1 Jan 2011 – 31 Dec 2011 (365 days)

# of cars

Total 100,581 95,386 95,293 88,345 128,789 114,942

Anomaly

4,259 2,475 4,365 3,891 6,089 5,603

Segment= 50-m length section

13

Parameter Estimation


Poisson mixture at a certain segment in the inbound Shibuya route

Estimated Poisson mixture fits the original histogram

Component Poisson multiplied by coefficient

14Incident DetectionResult – ROC curve

outbound (best) outbound (worst)


Baseline: [Zhu, et al. 2009]

better

Area Under the CurveProposed: 0.812Baseline: 0.794

15The Most Anomalous Trajectory


Accident

speed [km/h]

16

Outline



them




17

Discussion

• False Positives• Our definition of “incident” also matches abnormally fast

cars• Deeper analysis on alarmed trajectories is needed

• Proposed method can detect anomalies in real time• Learn traffic state model and compute divergence matrix in

advance• Use sliding window for each probe car

• Our approach performed better than an existing method based on physical traffic model• Congestion often occurs without any incidents on Shutoko;

there are many bottlenecks• Finding difference between current and usual traffic states

has an effect on traffic incident detection


18Road Characteristic Analysis Based on TSM


Fastest state

Slowest state

Usually slow section

19

Conclusions• Traffic state model

Mixing shared states in different ratio for each segment

• Detection methodCompare current/usual traffic states find irregular events

• Experimental resultBetter detection performance than existing method using real probe-car data in Shutoko

• Possible applications of traffic state modelTraffic characteristic analysis


・・・

G M C S G C M C S G

G M C S G G G G G GUsual

Cur.

20Some of Future Work(Working Now)

• (Automatic) Parameter tuning• K: the number of traffic states• N: the length of sliding window

• Proper divergence function selection• More deep analysis on anomalous

trajectories


21

Thank you for listening• Traffic state model

Mixing shared states in different ratio for each segment

• Detection methodCompare current/usual traffic states find irregular events

• Experimental resultBetter detection performance than existing method using real probe-car data in Shutoko

• Possible applications of traffic state modelTraffic characteristic analysis


・・・

G M C S G C M C S G

G M C S G G G G G GUsual

Cur.

Traffic Incident Detection Using Probabilistic Topic Model

Documents