You Are How You Drive: Peer and Temporal-Aware Representation - - PowerPoint PPT Presentation

Pengyang Wang, Yanjie Fu, Jiawei Zhang, Pengfei Wang, Yu Zheng, Charu Aggarwal

You Are How You Drive: Peer and Temporal-Aware Representation Learning for Driving Behavior Analysis

1

Outline

¨ Background and Motivation

¨ Definition and Problem Statement ¨ Methodology ¨ Application ¨ Evaluation ¨ Conclusion

Background and Motivation

¨ Car accident facts

2

Driving Behaviors

3

Driving Behaviors

4

It is essential to learn the pattern of driving behaviors

Challenges & Insights

¨ Challenge I: GPS traces – Non-applicable

GPS traces (e.g., time, latitude, longitude) encode the driving

• perations, states, and styles in a semantically implicit way

¨ Insight I:

Transforming GPS traces into graphs Convenient for representation learning

4

Challenges & Insights

¨ Challenge II: How to model dependencies?

peer dependencies temporal dependencies

¨ Insight II

jointly model the graph-graph peer dependency across drivers, as well as the current-past temporal dependency within a driver, in representation learning.

7

8

Outline

¨ Background and Motivation

¨ Definition and Problem Statement

¨ Methodology ¨ Application ¨ Evaluation ¨ Conclusion

Definition I

¨ Driving Operation

9

Driving operations are defined as a set of activities and steps that a driver operates when driving a vehicle, according to the driver’s personal judgment, experience and skills.

Speed-related: acceleration, deceleration, constant speed Direction-related: turning left, turning right, moving straight

Definition II

¨ Driving State

10

A driving state concerns the way that a vehicle moves at a specific time point or in a small time window. In other words, a driving state of a vehicle contains both the speed status (i.e., acceleration, deceleration, constant speed) and the direction status (i.e., turning left, turning right, moving straight) of a vehicle. For instance, a driving state example

• f a car can be <constant speed, moving straight>.

Definition III

¨ Driving State Transition Graph

11

Driving State 1 Driving State 2 Driving State 3 Driving State 4 Driving State 5

Problem Statement

¨ Given

• a driver (a vehicle)
• corresponding GPS trajectories

¨ Objective

• learning a mapping function

¨ Core tasks

• Constructing multi-view driving state transition graphs
• Automated profiling of driving behavior via peer and

temporal-aware representation learning

• Applications to transportation safety

11

f : D → V

D = [< t, ϕt, λt >]T

t=1

V = [vn]N

n=1

a sequence of time-varying yet relational vectorized representations

Framework Overview

12

• Driving State Transition Graph Sequence

Transition Probability View Transition Duration View Encoder

t=1

• t=2

t=3 t=T

Peer Dependency Representation Result (Vectors) Risky Area Detection Prediction and Historical Assessment of Driving Scores GPS Trajectory PTARL Decoder Temporal Dependency

……

13

Outline

¨ Background and Motivation ¨ Problem Statement

¨ Methodology ¨ Application

¨ Evaluation ¨ Conclusion

Methodology

¨ Construction of multi-view driving state transition

graphs

¨ Peer and temporal-aware representationlearning

14

Construction of multi-view driving state transition graphs

v Detecting Driving Operations v Extracting Driving State Sequences

15

v Detection of driving-related operations. v Detection of direction-related operations.

(1)acceleration while turning right, (2)acceleration while turning left, (3)acceleration while straightforward, (4)deceleration while turning right, (5)deceleration while turning left, (6)deceleration while straightforward, (7)constant speed while turning right, (8)constant speed while turning left, (9)constant speed while straightforward

Construction of multi-view driving state transition graphs

17

v Constructing Multi-view Driving State Transition Graphs

Driving State 1 Driving State 2 Driving State 3 Driving State 4 Driving State 5 Driving State 1 Driving State 2 Driving State 3 Driving State 4 Driving State 5

Transition probability view Transition duration view

Peer and temporal-aware representation learning

¨ Intuition 1: Structural Reservation ¨ Intuition 2: Temporal Dependency ¨ Intuition 3: Peer Dependency

18

For Intuition 1: Structural Reservation

¨ Base Model - Autoencoder

19

     y1

i

= σ(W1xi + b1), yk

i

= σ(Wkyk−1

i

+ bk), ∀k ∈ {2, 3, · · · , o}, zi = σ(Wo+1yo

i + bo+1).

     ˆ yo

i

= σ( ˆ Wo+1zi + ˆ bo+1), ˆ yk−1

i

= σ( ˆ Wkˆ yk

i + ˆ

bk), ∀k ∈ {2, 3, · · · , o}, ˆ xi = σ( ˆ W1ˆ y1

i + ˆ

b1).

… … … … … … …

x ˆ x

ˆ y1 y1

z

input matrix D d1 d2 dN

For Intuition 2: TemporalDependency

20

         #Sequential Encode Step (y1

i )τ

= σ(W1xτ

i + b1),

(yk

i )τ

= σ(Wk(yk−1

i

)τ + bk), ∀k ∈ {2, 3, · · · , o}, zτ

i

= (1 − cτ)zτ−1

i

+ cτ˜ zτ

i .

         #Sequential Decode Step (ˆ yo

i )τ

= σ( ˆ Wo+1zτ

i + ˆ

bo+1), (ˆ yk−1

i

)τ = σ( ˆ Wk(ˆ yk

i )τ + ˆ

bk), ∀k ∈ {2, 3, · · · , o}, ˆ xτ

i

= σ( ˆ W1(ˆ y1

i )τ + ˆ

b1).

For Intuition 3: Peer Dependency

21

Hc(Gτ) = X

ui2U

X

uj2U,ui6=uj

sτ

i,j · kzτ i zτ j k

2 2

the similarity of driving behavior between the driver 𝒗𝒋 and 𝒗𝒌 at the time slot 𝝊.

Objective Function

22

min 1 2 X

τ∈T

{ X

ui∈U(n)

k(xτ

i ˆ

xτ

i )k

2 2 + α · Hc(Gτ)}

Temporal Dependencies Peer Dependencies

13

Outline

¨ Background and Motivation ¨ Problem Statement ¨ Methodology

¨ Application

¨ Evaluation ¨ Conclusion

Application

¨ Prediction and Historical Assessment of Driving

Scores

¨ Risky Area Detection

24

25

Outline

¨ Background and Motivation ¨ Definition and Problem Statement ¨ Methodology ¨ Application

¨ Evaluation

¨ Conclusion and Future Work

Evaluation

26

¨ Data Description

From Beijing City Driving Score Distribution

Evaluation

27

¨ Baselines

(1) Auto-Encoder: minimizes the loss between the original feature representations and reconstructed ones. (2) DeepWalk: uses local information obtained from truncated random walks to learn latent representations. (3) LINE : optimizes the objective function that preserves both the local and global network structures with an edge-sampling algorithm. (4) Driving State Vector (DSV) : the traditional transportation approach.

Evaluation

29

¨ Evaluation Metrics

v Square Error

• Measure regression errors

v Coefficient of Determination

• measure the regression accuracy

v Normalized Discounted Cumulative Gain(NDCG@N)

• Evaluate the rankingperformance at TopN

v Kendall’s Tau Coefficient(Tau)

• Measure the overall ranking accuracy.

Overall performance

30

Square Error

0.4 0.6 0.8 1.0 1.2

Auto−Encoder DeepWAalk CNN LINE DSV PTARL

@5 @10 @15 @20

NDCG@N

0.0 0.2 0.4 0.6 0.8 1.0 1.2

Auto−Encoder DeepWAalk CNN LINE DSV PTARL R2

−1.0 −0.5 0.0 0.5

Auto−Encoder DeepWAalk CNN LINE DSV PTARL Tau

−0.1 0.0 0.1 0.2 0.3 0.4

Auto−Encoder DeepWAalk CNN LINE DSV PTARL

Robustness Check

31

Robustness check in the score-based group

Robustness Check

31

Robustness check in the driving-state-based group

Study of Peer and Temporal Dependencies

32

Square Error

0.2 0.4 0.6 0.8 1.0

Auto−Encoder PTARL−peer PTARL−temporal PTARL

@5 @10 @15 @20

NDCG@N

0.0 0.2 0.4 0.6 0.8 1.0 1.2

Auto−Encoder PTARL−peer PTARL−temporal PTARL Tau

−0.1 0.0 0.1 0.2 0.3 0.4

Auto−Encoder PTARL−peer PTARL−temporal PTARL R2

−1.0 −0.5 0.0 0.5

Auto−Encoder PTARL−peer PTARL−temporal PTARL

Study of Performance in Different Views

33

Square Error

0.2 0.3 0.4 0.5 0.6 0.7

Transition Probability View Transition Duration View PTARL R2

−1.0 −0.5 0.0 0.5

Transition Probability View Transition Duration View PTARL Tau

−0.1 0.0 0.1 0.2 0.3 0.4 0.5

Transition Probability View Transition Duration View PTARL

@5 @10 @15 @20

NDCG@N

0.0 0.2 0.4 0.6 0.8 1.0 1.2

Transition Probability View Transition Duration View PTARL

Historical Assessment of Driving Scores

34

5 10 15 20 0.0 0.2 0.4 0.6 0.8 1.0 1.2

Time Score

Riskier Driver Safer Driver

Risky Area Detection

35

t=1 t=2 t=3 t=4

35

Outline

¨ Background and Motivation ¨ Definition and Problem Statement ¨ Methodology ¨ Application ¨ Evaluation

¨ Conclusion

Conclusion

36

¨ We investigated driving behavior analysis from the

perspective of representation learning.

¨ We developed an analytic framework that jointly

modeled the peer and temporal dependencies

¨ constructing multi-view driving state transition graphs from

GPS traces to characterize driving behavior.

¨ incorporating the idea of gated recurrent unit to model both

the graph-graph peer dependency and integrating graph- graph peer penalties to capture the current-past temporal dependency in a unified optimization framework,

¨ applying our proposed method to enable the applications of

driving score prediction and risky area detection

¨ The method is effective.

38 Thanks!

Questions?