1 of 34

DF-TAR: A Deep Fusion Network for Citywide Traffic Accident Risk Prediction with Dangerous Driving Behavior

Patara Trirat (Presenter) and Jae-Gil Lee*

Data Mining Lab, School of Computing, KAIST

​

https://github.com/kaist-dmlab/DF-TAR

#TheWebConf | Paper No: 218

Image from KoROAD

2 of 34

DF-TAR: A Deep Fusion Network for Citywide Traffic Accident Risk Prediction with Dangerous Driving Behavior

​

#TheWebConf | Paper No: 218

Contents

  1. Introduction
  2. Preliminaries
  3. Methodology
    1. Correlation Analysis
    2. DF-TAR Model
  4. Experiments
  5. Conclusion

2

3 of 34

1. Introduction — Background

Traffic Accident

DF-TAR: A Deep Fusion Network for Citywide Traffic Accident Risk Prediction with Dangerous Driving Behavior

​

#TheWebConf | Paper No: 218

  • Fast urbanization
  • Crucial social-economic issue
  • About 50 million people are injured
  • Around 1.35 million people are killed
  • Mostly adolescents

3

4 of 34

1. Introduction — Background

Causes of Traffic Accident

DF-TAR: A Deep Fusion Network for Citywide Traffic Accident Risk Prediction with Dangerous Driving Behavior

​

#TheWebConf | Paper No: 218

4

Weather

Road Condition

Traffic Volume

Zoning

Time

Traffic Accident

5 of 34

1. Introduction — Related Work

Categorization of Previous Studies

DF-TAR: A Deep Fusion Network for Citywide Traffic Accident Risk Prediction with Dangerous Driving Behavior

​

#TheWebConf | Paper No: 218

5

1. Spatial Aspect

Grid

Actual Zone

2. Temporal Aspect

Daily

3. Learning Aspect

Classic Learning

Deep Learning

Hourly

Minutely

Input Features

Predictive Model

6 of 34

1. Introduction — Related Work

Our Approach

DF-TAR: A Deep Fusion Network for Citywide Traffic Accident Risk Prediction with Dangerous Driving Behavior

​

#TheWebConf | Paper No: 218

6

1. Spatial Aspect

Grid

Actual Zone

2. Temporal Aspect

Daily

3. Learning Aspect

Classic Learning

Deep Learning

Hourly

Minutely

?

Input Features

Predictive Model

7 of 34

1. Introduction — Motivation

Research Gap: Dangerous Driving Behavior

DF-TAR: A Deep Fusion Network for Citywide Traffic Accident Risk Prediction with Dangerous Driving Behavior

​

#TheWebConf | Paper No: 218

  • Dangerous Driving Behavior is a crucial factor that has not yet been utilized.
  • Also, dangerous driving is one of the most significant factors that causes traffic accidents.
  • Diagnosing dangerous driving behavior becomes feasible because it can be collected using various in-vehicle sensors or devices (e.g., Digital TachoGraph and On-Board Diagnostics).

7

Image from www.cheapfullcoverageautoinsurance.com

8 of 34

1. Introduction — Key Contributions

Research Questions

DF-TAR: A Deep Fusion Network for Citywide Traffic Accident Risk Prediction with Dangerous Driving Behavior

​

#TheWebConf | Paper No: 218

RQ1. Is the number of dangerous driving cases for each offense type correlated with the number of traffic accidents?

RQ2. Do the statistics of dangerous driving cases improve the performance of traffic accident risk prediction?

8

9 of 34

1. Introduction — Key Contributions

Overview of the Methodology

DF-TAR: A Deep Fusion Network for Citywide Traffic Accident Risk Prediction with Dangerous Driving Behavior

​

#TheWebConf | Paper No: 218

Examining the usefulness of dangerous driving statistics through geographical and temporal correlation analysis.

9

10 of 34

1. Introduction — Key Contributions

Overview of the Methodology

DF-TAR: A Deep Fusion Network for Citywide Traffic Accident Risk Prediction with Dangerous Driving Behavior

​

#TheWebConf | Paper No: 218

Proposing a deep fusion network, DF-TAR, to effectively incorporate dangerous driving statistics with other features and evaluating its performance through extensive experiments with an ablation study.

10

11 of 34

DF-TAR: A Deep Fusion Network for Citywide Traffic Accident Risk Prediction with Dangerous Driving Behavior

​

#TheWebConf | Paper No: 218

Contents

  • Introduction
  • Preliminaries
  • Methodology
    • Correlation Analysis
    • DF-TAR Model
  • Experiments
  • Conclusion

11

12 of 34

2. Preliminaries

Digital Tachograph (DTG) Data

DF-TAR: A Deep Fusion Network for Citywide Traffic Accident Risk Prediction with Dangerous Driving Behavior

​

#TheWebConf | Paper No: 218

  • DTG is a record-keeping instrument for driving log data
  • Two monthly datasets (Sept. 2016 & Sept. 2018) collected in five metropolitan cities (Seoul, Busan, Daejeon, Gwangju, and Daejeon) are provided.
  • Ten commercial vehicle types, taxis (personal and corporate), buses (town, city, rural, intercity, express, and rent), and trucks (personal and general), are included

12

Key features of a DTG record

13 of 34

2. Preliminaries

Dangerous Driving Behavior Criteria

DF-TAR: A Deep Fusion Network for Citywide Traffic Accident Risk Prediction with Dangerous Driving Behavior

​

#TheWebConf | Paper No: 218

There are nine types of dangerous driving offenses, (long-term) overspeed (OS), rapid acceleration (RA), quick start (QS), rapid deceleration (RD), sudden stop (SS), sudden lane change (SLC), sudden overtaking (SO), sharp turn (ST), and sudden u-turn (SUT).

13

14 of 34

2. Preliminaries

Traffic Accident Risk

DF-TAR: A Deep Fusion Network for Citywide Traffic Accident Risk Prediction with Dangerous Driving Behavior

​

#TheWebConf | Paper No: 218

14

Severity Level:

1 = slight injury with consciousness

2 = small injury without consciousness

3 = serious injury

4 = death

​

district

time interval

# of injured people of a severity level in a given district during a time interval

15 of 34

2. Preliminaries — Problem Statement

Traffic Accident Risk Prediction

DF-TAR: A Deep Fusion Network for Citywide Traffic Accident Risk Prediction with Dangerous Driving Behavior

​

#TheWebConf | Paper No: 218

Given

  • Historical traffic accident risk scores, environmental features, and dangerous driving cases

​

Predict

  • The future traffic accident risk scores

​

Objective

  • To minimize the prediction errors

15

16 of 34

DF-TAR: A Deep Fusion Network for Citywide Traffic Accident Risk Prediction with Dangerous Driving Behavior

​

#TheWebConf | Paper No: 218

Contents

  • Introduction
  • Preliminaries
  • Methodology
    • Correlation Analysis
    • DF-TAR Model
  • Experiments
  • Conclusion

16

17 of 34

3. Methodology — Correlation Analysis

Key Findings: A Counter-Intuitive Result

DF-TAR: A Deep Fusion Network for Citywide Traffic Accident Risk Prediction with Dangerous Driving Behavior

​

#TheWebConf | Paper No: 218

  • OverSpeed behavior not only shows the inconsistent correlation, but the computed scores are not statistically significant (p-value > .10)
  • It is the lowest ratio (~0.1%) of occurrences compared to other violations.
  • Potential reasons:
    • Mostly occurred at very late night
    • High-skilled drivers

17

18 of 34

3. Methodology — Correlation Analysis

Key Findings: Top-3 Dangerous Driving Offenses

DF-TAR: A Deep Fusion Network for Citywide Traffic Accident Risk Prediction with Dangerous Driving Behavior

​

#TheWebConf | Paper No: 218

Geographical Aspect (Subdistrict)

  1. Rapid Acceleration (0.88)
  2. Quick Start (0.84)
  3. Sharp Turn (0.83)

Temporal Aspect (Hour Interval)

  1. Quick Start (0.81)
  2. Sharp Turn (0.70)
  3. Sudden U-Turn (0.67)

18

Overall correlation scores

19 of 34

DF-TAR: A Deep Fusion Network for Citywide Traffic Accident Risk Prediction with Dangerous Driving Behavior

​

#TheWebConf | Paper No: 218

Contents

  • Introduction
  • Preliminaries
  • Methodology
    • Correlation Analysis
    • DF-TAR Model
  • Experiments
  • Conclusion

19

20 of 34

3. Methodology — DF-TAR Model

Model Architecture

DF-TAR: A Deep Fusion Network for Citywide Traffic Accident Risk Prediction with Dangerous Driving Behavior

​

#TheWebConf | Paper No: 218

20

21 of 34

3. Methodology — DF-TAR Model

Model Architecture: Convolutional Block

DF-TAR: A Deep Fusion Network for Citywide Traffic Accident Risk Prediction with Dangerous Driving Behavior

​

#TheWebConf | Paper No: 218

21

22 of 34

3. Methodology — DF-TAR Model

Model Architecture: Recurrent Block

DF-TAR: A Deep Fusion Network for Citywide Traffic Accident Risk Prediction with Dangerous Driving Behavior

​

#TheWebConf | Paper No: 218

22

23 of 34

3. Methodology — DF-TAR Model

Model Architecture: Fusion Block

DF-TAR: A Deep Fusion Network for Citywide Traffic Accident Risk Prediction with Dangerous Driving Behavior

​

#TheWebConf | Paper No: 218

23

24 of 34

3. Methodology — DF-TAR Model

Model Architecture: Fully-Connected Block

DF-TAR: A Deep Fusion Network for Citywide Traffic Accident Risk Prediction with Dangerous Driving Behavior

​

#TheWebConf | Paper No: 218

24

25 of 34

DF-TAR: A Deep Fusion Network for Citywide Traffic Accident Risk Prediction with Dangerous Driving Behavior

​

#TheWebConf | Paper No: 218

Contents

  • Introduction
  • Preliminaries
  • Methodology
    • Correlation Analysis
    • DF-TAR Model
  • Experiments
  • Conclusion

25

26 of 34

4. Experiments — Experimental Settings

Data Sets

DF-TAR: A Deep Fusion Network for Citywide Traffic Accident Risk Prediction with Dangerous Driving Behavior

​

#TheWebConf | Paper No: 218

  • Traffic Accident Risk Scores (1 features)
  • Dangerous Driving Behavior (9 features)
  • Static Environmental Features (98 features)
    • Demographic Data
    • Point-of-Interest Data
    • Road Network and Specification
  • Dynamic Environmental Features (24 features)
    • Weather and Air Quality Data
    • Traffic Volume
    • Calendar Data

26

Train (70%): Sept. 1 - 21

Test (30%): Sept. 22 - 30

27 of 34

4. Experiments — Experimental Settings

Comparison Study

DF-TAR: A Deep Fusion Network for Citywide Traffic Accident Risk Prediction with Dangerous Driving Behavior

​

#TheWebConf | Paper No: 218

  • Historical Average (HA)
  • Linear Regression (LR)
  • eXtreme Gradient Boosting (XGB)
  • SDAE
  • TARPML
  • Hetero-ConvLSTM
  • TA-STAN

27

Baselines

Evaluation Metrics

28 of 34

4. Experiments — Experimental Settings

Hyperparameter Settings

DF-TAR: A Deep Fusion Network for Citywide Traffic Accident Risk Prediction with Dangerous Driving Behavior

​

#TheWebConf | Paper No: 218

28

Input sequence length

12 hours

Output sequence length

12 hours

Standardization

Min-Max Normalization

Optimization

Adam optimizer

Learning Rate

Initial with 0.001 with a decaying technique

Batch Size

32

Early Stopping

Applied

29 of 34

4. Experiments — Results

Overall Results

DF-TAR: A Deep Fusion Network for Citywide Traffic Accident Risk Prediction with Dangerous Driving Behavior

​

#TheWebConf | Paper No: 218

29

Sept. 2016

Sept. 2018

30 of 34

4. Experiments — Results

Ablation Study

DF-TAR: A Deep Fusion Network for Citywide Traffic Accident Risk Prediction with Dangerous Driving Behavior

​

#TheWebConf | Paper No: 218

30

Model performance with the different feature set

Improvement of up to 32%

Improvement of up to 5%

31 of 34

4. Experiments — Results

Case Study

DF-TAR: A Deep Fusion Network for Citywide Traffic Accident Risk Prediction with Dangerous Driving Behavior

​

#TheWebConf | Paper No: 218

31

32 of 34

DF-TAR: A Deep Fusion Network for Citywide Traffic Accident Risk Prediction with Dangerous Driving Behavior

​

#TheWebConf | Paper No: 218

Contents

  • Introduction
  • Preliminaries
  • Methodology
    • Correlation Analysis
    • DF-TAR Model
  • Experiments
  • Conclusion

32

33 of 34

5. Conclusion

Conclusion and Future Work

DF-TAR: A Deep Fusion Network for Citywide Traffic Accident Risk Prediction with Dangerous Driving Behavior

​

#TheWebConf | Paper No: 218

  • We propose DF-TAR model by incorporating the dangerous driving statistics.
  • We also quantify the correlation scores between the dangerous driving behavior and the past accidents to see whether they are strongly correlated.
  • The evaluation results show that the our model trained with dangerous driving behavior improved up to 54% in MAE and 18% in RMSE.
  • For future work, dangerous driving behavior can be utilized with real-time traffic accident prediction systems to help individuals avoid potential risks associated with each vehicle on the road.

33

34 of 34

Thank You!

DF-TAR: A Deep Fusion Network for Citywide Traffic Accident Risk Prediction with Dangerous Driving Behavior

Patara Trirat (Presenter) and Jae-Gil Lee*

Data Mining Lab, School of Computing, KAIST

#TheWebConf | Paper No: 218

Icon Credit: flaticon.com

https://github.com/kaist-dmlab/DF-TAR