Learning Deep Representation from Big and Heterogeneous Data for Traffic Accident Inference
1. 1 KYOTO UNIVERSITY
KYOTO UNIVERSITY
Learning Deep Representation from Big and
Heterogeneous Data for Traffic Accident Inference
Daiki Tanaka
Kashima lab., Kyoto University
Paper Reading Seminar, 2017/10/13(Fri)
2. 2 KYOTO UNIVERSITY
Today’s paper:
n Title : Learning Deep Representation from Big and
Heterogeneous Data for Traffic Accident Inference
(AAAI’16)
n Authors:
Quanjun Chen Xuan Song Harutoshi Yamada Ryosuke Shibasaki
Center for Spatial Information Science, The University of Tokyo
4. 4 KYOTO UNIVERSITY
Background:
n The increasing number of transportation vehicles causes problems
such as traffic jams and traffic accidents.
n Some of them such as traffic jams are alleviated, by using the real-
time traffic volume data and vehicle navigation systems based on
GPS.
n But traffic accidents still need to be solved.
n Understanding what causes traffic accident is crucial.
5. 5 KYOTO UNIVERSITY
Problem setting:
n “Can we estimate traffic accident risk just as traffic jam through
real-time location data?”
n It is difficult to predict traffic accident or not. Because traffic
accidents are caused by complex factors.
p Input : human mobility and traffic accident data
p Output : prediction of risk level for regions
7. 7 KYOTO UNIVERSITY
Proposed method:
data used in this paper
n Traffic accident data (300,000 records)
l occurrence location
l hourly occurrence time
l Severity level
• severity is graded as three levels.
n Human mobility data
l GPS record of 1,600,000 people (2013/1/1〜2013/7/31)
• GPS information is uploaded every 5 minutes.
1 2 3
9. 9 KYOTO UNIVERSITY
Proposed method:
preprocessing on dataset
n We mesh map into 500m × 500m square.
n We select one hour as the time interval.
n We define risk level 𝑔",$ as follows :
l If traffic accident happened 𝑛 times in region 𝑟 at time 𝑡,
𝑔",$ = ) 𝑆+,",$
,
+-.
l 𝑆+,",$ is the severity of 𝑖 -th traffic accident.
10. 10 KYOTO UNIVERSITY
Proposed method:
human mobility
n We define 𝑑",$ as the mean density of GPS records in region
r and time t of different days.
• Human mobility follows a stationary pattern except some special
days.
n Human mobility matrix 𝒅2,",$ with size (2m+1) × (2m+1)
and centered on region r, should be used instead of single
region.
• Traffic accident risk in a region may be affected by human mobility
of neighbor regions.
12. 12 KYOTO UNIVERSITY
Proposed method:
AutoEncoder
n Autoencoder is single-layer network and trained in un-supervised
way.
n Input : a set of x
• 𝒛 = 𝑠 𝑾𝒙 + 𝒃 : encode
• 𝒚 = 𝑠 𝑾? 𝒛 + 𝒃? : decode
(s is non-linear function.)
n Model parameters 𝜽 are optimized by
minimizing error 𝐿(𝒙, 𝒚) as:
• 𝜽 = argminU 𝐿(𝒙, 𝒚)
13. 13 KYOTO UNIVERSITY
Proposed method:
denoise AutoEncoder
n Denoise autoencoder is based on autoencoder.
n Train samples are added into noise.
n Model parameters 𝜽 are optimized by
minimizing error 𝐿(𝒙′, 𝒚) as:
• 𝜽 = argmin
U
𝐿(𝒙′, 𝒚)
• 𝒙′ = 𝒙 + 𝛿(noise)
l Denoise autoencoder has an ability to remove noise.
16. 16 KYOTO UNIVERSITY
experiment:
setting
n How to use data
l 80% : using for the model training
l 20% : using for testing and evaluation
n SdAE architecture parameters
l There are three denoise autoencoder layers.
l The number of units in each layer is [40, 40, 40].
20. 20 KYOTO UNIVERSITY
n They investigated how human mobility affects traffic accident risk.
n They have utilized a deep architecture to extract features from
human mobility data.
n They trained a prediction model for simulating accident risk on large
scale and in real-time.
n They will combine human mobility with other data like land uses and
Points Of Interest data to improve their model.
Conclusion and future works: