Driving Behavior Analysis and Trafﬁc Improvement using Onboard

Sensor Data and Geographic Information

Jun-Zhi Zhang and Huei-Yung Lin

Department of Electrical Engineering, National Chung Cheng University, Chiayi 621, Taiwan

Keywords:

Driving Behavior Analysis, Geographic Information System, Data Mining.

Abstract:

In this paper, we present a method to extract the training and testing data from geographic information system

(GIS) and global position system (GPS) for neural networks. Trafﬁc signs, trafﬁc lights and road information

from the OpenStreetMap (OSM) and the government platform are compared with driving data and videos

to extract images containing the important information. We also propose trafﬁc improvement suggestions

for intersections or roads by analyzing the relationship between driving behaviors, trafﬁc lights, and road

infrastructures. We use OBD-II and CAN bus logger to record more driving information, such as engine speed,

vehicle speed, steering wheel steering angle, etc. We analyze the driving behavior using sparse automatic

encoders and data exploration to detect abnormal and aggressive behavior. The relationship between the

aggressive driving behavior and road facilities is derived by regression analysis, and some suggestions are

provided for improving speciﬁc intersections or roads.

1 INTRODUCTION

In recent years, the ﬁeld of self-driving cars has be-

come more and more popular. Many companies and

research institutes have started the development of au-

tonomous vehicle systems. It is very likely to have

self-driving cars and the vehicles controlled by human

drivers co-exist on the road in the future. In this re-

gard, one of the key components to the success of self-

driving systems is to understand the human driving

behavior in order to avoid the human-machine con-

ﬂict (Dong and Lin, 2021). With the recent advances

of machine learning techniques, the data-driven ap-

proaches have made the complicated human behavior

modeling move a big step forward. Especially, some

researches with signiﬁcant progress have been con-

ducted using deep learning (Hartford et al., 2016).

For the modern learning approaches, in addition to

the design of network structures, another major issue

is the requirement of a large amount of training data.

In the automative applications, this usually involves

the data collection from on-board sensors and the in-

formation extraction for speciﬁc analysis purposes.

These might include the images captured by the in-

car cameras for environment perception, and the pro-

prioceptive driving data recorded by the on-board di-

agnostics systems. In either case, it is necessary to

https://orcid.org/0000-0002-6476-6625

extract proper data segments for neural network train-

ing and testing. For instance, learning the road sign

recognition uses certain trafﬁc scene images, or mod-

eling the driver’s acceleration behavior uses selected

gas pedal information. The use of large datasets for

training is commonly agreed for deep neural networks

to perform well or better.

In the early stage of related research, data annota-

tion or labeling are mostly done manually, and some-

times through crowdsourcing such as using Amazon

Mechanical Turk. For driving images, the dataset col-

lected according to different tasks contains a variety

of scenes and features. Since the selection and ﬁl-

tering of adequate data require signiﬁcant time and

human labor, it motivates a data management prob-

lem: How to search speciﬁc trafﬁc scenes within a

large amount of image sequences? In this work, we

present a road scene extraction system for speciﬁc

landmarks and indicators of the transportation infras-

tructure. The information derived from GIS (geo-

graphic information system) and GPS are used with

the recorded driving videos to identify the road scenes

with static objects such as trafﬁc lights, trafﬁc signs,

bridge and tunnel, etc.

On the other hand, in addition to the exterocep-

tive sensors (such as LiDAR, GPS, camera, etc.), the

information collected from prioceptive sensors of the

vehicle can also be used to analyze the driving behav-

284

Zhang, J. and Lin, H.

Driving Behavior Analysis and Trafﬁc Improvement using Onboard Sensor Data and Geographic Information.

DOI: 10.5220/0010384102840291

In Proceedings of the 7th International Conference on Vehicle Technology and Intelligent Transport Systems (VEHITS 2021), pages 284-291

ISBN: 978-989-758-513-5

ior (Yeh et al., 2019). The sensor data derived directly

from the vehicle operation can provide more compre-

hensive driving information. It allows the researchers

and practitioners to study driving behaviors and trafﬁc

safety issues more precisely. In this paper, we adopt

OBD-II (on-board diagnostics) and CAN bus (Con-

troller Area Network) logger to collect data. By mea-

suring a number of parameters at high sampling rate,

it is possible to fully observe the driving behaviors in

real life, and understand how they are affected by the

trafﬁc and road infrastructures.

To analyze the relationship between the driving

behavior and the transportation infrastructure, its vi-

sualization on the map provides a way for better ob-

servation and investigation. We use machine learning-

based methods to extract the unique features from the

driving data and then map to the RGB color space to

visualize the driving behavior. The data mining algo-

rithms are adopted for data analysis and classify the

driving behavior into four categories, from normal to

aggressive. A regression analysis is then conducted

on the relationship between the aggressive driving be-

havior and the road features of intersections.

2 RELATED WORK

2.1 Training and Testing Data

Extraction

Due to the popularity of learning based algorithms in

recent years, the acquisition of training and testing

data has become an important problem. The current

data extraction approaches are mainly divided into

two categories, information-based and image-based

extraction. Hornauer et al. and Wu et al. (Hornauer

et al., 2019; Wu et al., 2018) proposed unsupervised

image classiﬁcation methods to extract to the images

similarly to those provided by general users (Zhirong

et al., 2018). In supervised classiﬁcation, the data are

labeled manually. People need to have similar under-

standing to annotate the same scene. Their network

presents a concept based on feature similarity for ﬁrst-

person driving image query. However, it is not satis-

factory for the requests of most users.

In addition to the image extraction and classiﬁ-

cation, Naito et al. developed a browsing and re-

trieval system for driving data analysis (Naito et al.,

2010). The system provides a multi-data browser, a

retrieval function based on query and similarity, and

a quick browsing function to skip extra scenes. For

the scene retrieval, the top N images highly similar

to the currently driving scenario are retrieved from

the database. In this technique, while the image se-

quence is processed, the system calculates the simi-

larity between the input scene and the scenes stored

in the database. A pre-deﬁned threshold is used to

identify the similarity between the images. Since the

method mainly searches the driving video itself, it is

not able to know if the images contain the objects or

information interested to the users for precise extrac-

tion.

2.2 Driving Behavior Analysis

In the past few years, the key technologies of automa-

tive driving assistance systems have become more

mature (Lin et al., 2020). However, the ‘autonomous’

vehicles are still not ready without the human drivers.

Due to the current limitations of driving assistance

systems, researchers and developers are seeking for

the solutions to enhance the human driving capability.

Since the driving habits are very difﬁcult to change, it

is expected to have a human-centered driving environ-

ment to avoid dangerous situations. By understand-

ing the relationship among the trafﬁc lights, road in-

frastructure and driving behavior, some transportation

improvement suggestions can be provided. Besides,

knowing the human reaction is also a crucial issue in

the future world with mixed human drivers and self-

driving cars.

For driving behavior analysis, Liu et al. proposed

a method using various types of sensors connected to

the control area network (Liu et al., 2014; Liu et al.,

2017). A deep sparse autoencoder is then used to ex-

tract the hidden features from driving data to visualize

the driving behavior. Alternatively, Constantinescu et

al. used both PCA and HCA methods to analyze the

driving data (Constantinescu et al., 2010). The perfor-

mance of the algorithms is veriﬁed by classifying the

driving behavior into six categories according to dif-

ferent aggressiveness. In the study of Kharrazi et al.,

the driving behavior is classiﬁed into three categories,

calm, normal and aggressive, by a method using quar-

tile and Kmeans (Kharrazi et al., 2019). The analysis

has demonstrated that Kmeans is able to provide good

driving behavior classiﬁcation results.

In the above methods, the correlation between the

driving behavior and the environment is not investi-

gated. For the discussion of more speciﬁc events, Tay

et al. used the regression model to associate driv-

ing accidents with the environment (Tay et al., 2008).

Wong et al. used a negative binomial regression to

analyze the number of driving accidents and the road

features of the intersection (Wong, 2019). It can help

us understand the relationship between the accidents

and road features. The road intersection can also be

Driving Behavior Analysis and Trafﬁc Improvement using Onboard Sensor Data and Geographic Information

285

Figure 1: The trafﬁc light information shown on the Open-

StreetMap. The yellow dots indicate the locations of trafﬁc

lights on main roads.

improved by the simulation carried out based on the

analysis results. Schorr et al. recorded the driving

data in one and two-way lanes (Schorr et al., 2016).

Based on ANOVA analysis, the conclusion about the

impact of the lane width to the driving behavior is

drawn. Mohammad et al. investigated how the acci-

dents were affected speciﬁc driving behaviors through

a number of questionnaires and interviews (Aboja-

radeh et al., 2014). They used regression analysis

to derive the correlation between the number of ac-

cidents and the types of dangerous driving behaviors.

Regarding the improvement on transportation in-

frastructures, various suggestions were proposed for

different designs of roads and intersections. Chunhui

et al. proposed to optimize the signal lights at the

intersections to make the pedestrian crossing easier

(Chunhui et al., 2017; Wang et al., 2019). The ef-

ﬁciency of intersections is improved by reducing the

conﬂicts between the turning vehicles and pedestri-

ans. Ma et al. proposed to add a dedicated left-turn

lane and left turn waiting area according to the aver-

age daily trafﬁc volume at the intersection (Ma et al.,

2017). The proposed method is able to accommodate

more vehicles waiting for left turn. They also an-

alyze three common left-turn operation scenarios at

the intersections and compare their differences. In

addition to the suggestions for road infrastructures,

there also exist some improvements based on the traf-

ﬁc light analysis. In the recent work, Anjana et al.

presented a method based on different trafﬁc volumes

at the intersections to evaluate the safety caused by the

green time of the trafﬁc light (Anjana and Anjaneyulu,

2015).

3 DATASET EXTRACTION

We ﬁrst collect trafﬁc lights, trafﬁc signs and road in-

formation on the OpenStreetMap (OSM) and the gov-

ernment’s GIS-T transportation geographic informa-

tion storage platform of the as the locations of inter-

est for image data extraction. Figure 1 illustrates an

example of the trafﬁc light information shown on the

OSM. The yellow dots indicate the locations of trafﬁc

lights on main roads. For image data extraction, the

transportation infrastructure and road information are

used to identify the locations of interest using the GPS

coordinates. We compare the GPS information of the

driving data and the locations of interest. The asso-

ciated images are then extracted and stored in video

sequences for speciﬁc application uses (such as the

training and testing data for trafﬁc light detection).

The speciﬁcations of the driving recorder contain

the images with the resolution of 1280 ×720 and 110

◦

FOV (ﬁeld-of-view) in the horizontal direction. To

extract the suitable image data, the users need to con-

sider a geographic range of the interested target. As

a typical example of road scene extraction with trafﬁc

lights, the size of the trafﬁc signal in the image might

be larger than 25 × 25 pixels for speciﬁc tasks. This

corresponds to about 50 meters away from the vehi-

cle, so the video should be pushed back 5 seconds to

start the image extraction.

A program interface is created for users to eas-

ily operate the data and assign the parameters. As

shown in Figure 2, it consists of folder selection, item

menu for extraction, OSM map display and driving

image screen. The user ﬁrst select the folder where

the driving record video and driving GPS information

are located, and the folder where the extracted image

will be stored, followed by the selection of the trafﬁc

infrastructure or road information to be extracted. On

the interface, the vehicle’s GPS trajectory and the user

selected trafﬁc infrastructures simultaneously overlay

on the OSM window, and the synchronized driving

video is displayed on the right for inspection.

4 DRIVING BEHAVIOR

ANALYSIS

In this paper, we mainly focus on the analysis of driv-

ing behavior and the correlation with trafﬁc and road

features. The common relationship is ﬁrst established

and the studies on speciﬁc scenarios are then carried

out. The driving behavior is classiﬁed into normal and

aggressive, and analyzed through data visualization

and the regression model on the number of aggressive

driving behaviors and road features.

VEHITS 2021 - 7th International Conference on Vehicle Technology and Intelligent Transport Systems

286

Figure 2: The user interface for image data extraction. 1.

folder selection, 2. processing status display, 3. extraction

item menu, 4. driving trajectory display, and 5. driving

video display.

4.1 Data Collection

The tools for data collection in this work include

ODB-II (Malekian et al., 2014) and CANbus loggers.

Unlike most previous work which only use the infor-

mation obtained from GPS receivers (with GPS mes-

sages, vehicle speed and acceleration), OBD-II and

CANbus loggers are able to collect various types of

driving data to analyze in more details. The speciﬁc

data types used for our driving behavior analysis are

as follows:

• OBD-II: engine rotating speed, engine load, throt-

tle pedal position, acceleration XYZ, and vehicle

speed.

• CANbus logger: engine rotating speed, throttle

pedal position, braking pedal position, steering

angle, wheel speed, and vehicle speed.

• GPS receiver: GPS and UTC.

In addition, we also use two public datasets, DDD17

dataset (Binas et al., 2017) and UAH Drive-set

(Romera et al., 2016). The datasets are acquired by

the driving monitoring application, DriveSafe, and

mainly used to verify the classiﬁcation and analysis

methods (Bergasa et al., 2014).

4.2 Visualization of Driving Behavior

The relationship between the driving behavior and

trafﬁc infrastructures can be observed by the data vi-

sualization on the map. We use the sparse autoen-

coder (SAE) to extract features from the driving data,

compress the high-dimensional features into three di-

mensions, and mapping to RGB space for display on

the OpenStreetMap (Liu et al., 2017). The loss func-

Figure 3: The ﬂowchart and network structure for driving

behavior analysis.

tion with sparse constraints is given by

sparse

(W, b) = J (W, b) + β

∑

j=1

KL(ρ k

) (1)

The difference between SAE and autoencoder (AE) is

that a penalty term is added to the loss function, so

the activation of the hidden nodes drops to the value

we need. Using this property, the relative entropy is

added to the loss function to penalize the value of the

average activation degree far away from the level ρ.

The parameters can keep the average activation de-

gree of hidden nodes at the level. Thus, the loss func-

tion only needs to add the penalty term of relative en-

tropy without sparse constraints.

Figure 3 illustrates the structure to visualize the

driving behavior. The network contains 9 hidden lay-

ers, and the dimensionality reduction of each layer is

half the number of nodes in the previous layer. Our

data collected by OBD-II contain 7 types, and become

70 dimensions after windowing process. Thus, the di-

mension reduction in the network is 70 → 35 → 17 →

8 → 3 → 8 → 17 → 35 → 70, and the features are

extracted by the last 5 layers. The data collected by

CANbus logger contain 6 types, and are processed to

60 dimensions after windowing. Likewise, the input

to the network consists of 60 nodes, and the dimen-

sion reduction is given by 60 → 30 → 15 → 7 → 3 →

7 → 15 → 30 → 60. Finally, the driving behavior is

visualized on the OpenStreetMap.

We use the Kmeans clustering algorithm to fur-

ther classify the driving behavior. The elbow method

is used to ﬁnd the most appropriate k value to clas-

sify the driving behavior according to different ag-

gressiveness (Thorndike, 1953). From normal to ag-

gressive, it is classiﬁed into four levels, and the most

aggressive driving behavior is marked on the OSM.

4.3 Negative Binomial Regression

We refer to (Wong, 2019) and use negative binomial

regression model to analyze the road features at in-

tersections and interchanges. It is an extended ver-

sion of Poisson regression to deal with the data over-

dispersed problem. The negative binomial regression

Driving Behavior Analysis and Trafﬁc Improvement using Onboard Sensor Data and Geographic Information

287

model

= exp(β

+ β

+ ··· + β

+ ε

) (2)

is used to predict the number of aggressive driving

behavior µ

, where β is the correlation term associated

with each road feature parameter, and ε

is an error

term. Next, we need to verify if the data are over-

dispersed, so Pearson’s chi-squared test is carried out

(Pearson, 1900). When the ratio is greater than 1, the

data is considered to be over-dispersed.

To evaluate whether the Poisson regression or

negative binomial regression can better ﬁt our data,

Akaike information criterion (AIC) is calculated for

these two models (Akaike, 1974). AIC is an effec-

tiveness measure of data ﬁtting on regression models

given by

AIC = 2k − 2ln(L) (3)

where k is the number of features and ln(L) is the

maximum likelihood. A smaller AIC value implies be

a better ﬁtting model. As an example case, the max-

imum likelihoods of Poisson and negative binomial

regression are -21.457 and -21.758 respectively, and

the AIC values are 58.914 and 59.516 respectively.

It shows that the negative binomial regression model

has a smaller AIC. Thus, it is used as the model for

our analysis.

After classifying the driving behavior by Kmeans,

it is found that the aggressive driving behavior occurs

more frequently at the interchanges and intersections.

The negative binomial regression analysis is carried

out on these two speciﬁc driving scenarios. We adopt

the road features proposed by Wong (Wong, 2019)

and those commonly appeared in Taiwan road scenes

as follows.

1. Interchange: (1) section length, (2) lane width, (3)

speed limit, (4) trafﬁc ﬂow.

2. 4-Arm Intersection: (1) without lane marking, (2)

straight lane marking, (3) left lane marking, (4)

right lane marking, (5) shared lane marking, (6)

shared lane marking at roadside, (7) motorcycle

priority, (8) branch road.

3. 3-Arm Intersection: (1) without lane marking, (2)

straight lane marking, (3) shared lane marking at

roadside, (4) lane ratio, (5) motorcycle priority,

(6) branch road.

5 EXPERIMENTS

The experiments contain two parts: One is the system

for the extraction of training and testing dataset, and

the other is the driving behavior analysis based on the

driving and road features.

Figure 4: The driving trajectory (marked in red curve) and

trafﬁc light positions (marked in purple circles) displayed

on the OpenStreetMap. The orange dots (a) and (b) corre-

spond to the images shown in Figures 5(a) and 5(b), respec-

tively.

(a) Long range image. (b) Short range image.

Figure 5: The images containing trafﬁc lights extracted

from the map in 4, corresponding to the locations (a) and

(b), respectively.

5.1 Extraction of Training and Testing

Data

In this experiment, we demonstrate the image data ex-

traction for the road scenes with trafﬁc lights. Fig-

ure 4 shows the driving trajectory (marked in red

curve) and trafﬁc light positions (marked in purple

circles) displayed on OSM. The driving video is ﬁl-

tered through the extraction system to contain the traf-

ﬁc lights from the far to near distance. The extracted

images as shown in Figures 5(a) and 5(b) correspond

to the orange dots (a) and (b) on the map (in Figure

4), respectively.

5.2 Driving Behavior Analysis

For driving behavior analysis, the visualization and

Kmeans classiﬁcation are presented ﬁrst, followed by

the analysis on the driving behavior and road features.

VEHITS 2021 - 7th International Conference on Vehicle Technology and Intelligent Transport Systems

288

Figure 6: The visualized driving behavior. The aggres-

sive driving data and trafﬁc lights are marked on the Open-

StreetMap. The red circle location corresponds to the data

enclosed by red in Figure 7 and the driving image shown in

Figure 8.

5.2.1 Visualization and Kmeans Classiﬁcation

In this experiment, whether Kmeans can effectively

classify the driving behavior is veriﬁed ﬁrst. We

use ﬁve segments of driving data in UAH Drive-set,

and the drivers are in normal and aggressive behav-

iors separately. In each data segment, 50 samples

are taken for classiﬁcation. The results are shown in

Table 1 with the percentage of correct classiﬁcation,

where D1 – D5 represent ﬁve drivers. N and A are

the normal and aggressive driving, respectively. The

table illustrates that Kmeans is able to provide satis-

factory classiﬁcation results on normal and aggressive

driving behaviors.

Figure 6 shows the driving behavior (including ag-

gressive driving) using the driving data collected by

ourselves visualized on OpenStreetMap with the traf-

ﬁc light location information. The driving data chart

and the image acquired by a car digital video recorder

are shown in Figures 7 and 8, respectively. By ob-

serving the information from these three aspects, the

correlation among them can be analyzed. In this ex-

ample, the aggressive driving behavior at location in-

dicated by the red circle (Figure 6) is caused by a

pedestrian passing through the intersection (Figure 8),

which leads to the braking and turning of the vehicle

(Figure 7).

By visualizing the driving behavior and displaying

the aggressive driving behaviors on OSM with refer-

ence to the driving video, we are able to observe the

correlation between the driving behavior and trafﬁc

infrastructure. Three situations are analyzed as fol-

lows.

a. The inﬂuence of two-way lanes on the driving be-

havior: We found that the vehicle speed in a two-

way lane is higher than a one-way lane. Thus,

Figure 7: The driving data chart. The red frame corresponds

to the location indicated by the red circle in Figure 6 and the

driving image in Figure 8.

Figure 8: The image acquired at the location corresponding

to the circle in Figure 6. A pedestrian passing through the

intersection leads the braking of the vehicle as illustrated in

Figure 7.

the aggressive driving behaviors with fast driving

and emergency braking are more likely to occur

in two-way lanes.

b. The inﬂuence of trafﬁc lights on the driving be-

havior: We found that most of the aggressive driv-

ing behaviors occurred at intersections. There

might be many reasons, such as fast changing sig-

nals and the poor design of the road. These gener-

ally cause more conﬂicts between the drivers and

other vehicles.

c. The inﬂuence of interchanges on the driving be-

havior: In the highway trafﬁc, we found that most

of the aggressive driving behaviors occur at inter-

changes. A vehicle entering the entrance of the

interchange tends to drive into the inner lane. This

generally causes the other drivers to change lanes

or slow down.

5.2.2 Negative Binomial Regression

Since the aggressive driving behaviors frequently oc-

cur near the intersections and interchanges, we further

investigate these driving scenarios using negative bi-

nomial regression analysis on the correlation between

the number of aggressive behaviors and road features.

Driving Behavior Analysis and Trafﬁc Improvement using Onboard Sensor Data and Geographic Information

289

Table 1: The Kmeans classiﬁcation performance on UAH Drive-set.

D1 D2 D3 D4 D5

N A N A N A N A N A

100% 80% 100% 100% 100% 96% 98% 100% 98% 98%

(a) The regression result at a 4-arms intersection.

(b) The regression result at a highway interchange.

Figure 9: The negative binomial regression analysis results

for a 4-arm intersection and a highway interchange. (a) In-

tercept: the error term of regression model, LEFT: left turn

lane mark, STRA: straight lane mark, RIGHT: right turn

lane mark, TWO: shared lane mark, SHARE: shared lane

mark on the side of the road, NO: no lane mark, MOTOR:

the number of priority locomotive lanes, CROSS: the num-

ber of branch roads and the coefﬁcient term is the param-

eter by the regression model. (b) LONG: the length of in-

terchange, LANE: the width of the lane, LIMIT: the ramp

speed limit, and FLOW: the average daily trafﬁc volume.

The P-value is used to evaluate whether the feature

has a signiﬁcant impact on aggressive driving behav-

ior (Dahiru, 2008). Two driving scenarios are exam-

ined as follows.

4-Arms Intersection: There are eight different road

features at the intersections as deﬁned previ-

ously. After the regression analysis as shown in

Figure9(a), we see the features that have great im-

pacts on the aggressive driving behaviors include

‘straight lane marking’, ‘shared lane marking at

roadside’ and ‘without lane marking’. The inﬂu-

ences of these features on the driving behaviors

are positive correlation, negative correlation and

positive correlation, respectively. When “P > |z|”

< 0.05, the feature is important affects the aggres-

sive behavior.

Highway Interchange: There are four different road

features deﬁned for the highway. After the regres-

sion analysis as shown in Figure 9(b), we see the

features that have great impacts on the aggressive

driving behaviors are ‘speed limit’ and ‘length of

interchange’. The inﬂuences of these features on

the driving behaviors are positive correlation and

negative correlation, respectively.

6 CONCLUSIONS

This paper presents the image data extraction based

on geographic information and driving behavior anal-

ysis using various types of driving data. The trafﬁc in-

frastructure and GPS information are used to extract

speciﬁc road scenes for network training and testing

purposes. We use OBD-II and CANbus loggers to

acquire driving data, and classify the driving behav-

iors using SAE feature extraction and Kmeans algo-

rithm. The negative binomial regression analysis is

performed for speciﬁc scenarios. Our result show that

lane ratios, without lane markings, and straight lane

markings are important features which affect the ag-

gressive driving behavior. In the end, we present the

trafﬁc improvements based on the analysis for a case

study at an intersection.

ACKNOWLEDGMENTS

This work was ﬁnancially/partially supported by Cre-

ate Electronic Optical Co., LTD, Taiwan.

REFERENCES

Abojaradeh, M., Jrew, B., and Al-Ababsah, H. (2014). The

effect of driver behavior mistakes on trafﬁc safety.

Journal of Civil and Envivornment Research, 6:39–

54.

Akaike, H. (1974). A new look at the statistical model iden-

tiﬁcation. IEEE Transactions on Automatic Control,

19:716—-723.

VEHITS 2021 - 7th International Conference on Vehicle Technology and Intelligent Transport Systems

290

Anjana, S. and Anjaneyulu, M. (2015). Safety analysis

of urban signalized intersections under mixed trafﬁc.

Journal of Safety Research, 52:9–14.

Bergasa, L. M., Almer

ıa, D., Almaz

an, J., Yebes, J. J., and

Arroyo, R. (2014). Drivesafe: an app for alerting inat-

tentive drivers and scoring driving behaviors. IEEE

Intelligent Vehicles Symposium(IV), pages 240–245.

Binas, J., Neil, D., Liu, S.-C., and Delbruck, T. (2017).

Ddd17: End-to-end davis driving dataset. Workshop

on Machine Learning for Autonomous Vehicles, pages

1–9.

Chunhui, Y., Wanjing, M., Ke, H., and Xiaoguang, Y.

(2017). Optimization of vehicle and pedestrian sig-

nals at isolated intersections. Transportation Research

Part B: Methodological, 98(C):135–153.

Constantinescu, Z., Marinoiu, C., and Vladoiu, M. (2010).

Driving style analysis using data mining techniques.

International Journal of Computers and Communica-

tions and Control, 5(5):654–663.

Dahiru, T. (2008). P-value, a true test of statistical signiﬁ-

cance? a cautionary note. Annals of Ibadan Postgrad-

uate Medicine, 6:21—-26.

Dong, B. T. and Lin, H. Y. (2021). An on-board monitoring

system for driving fatigue and distraction detection.

In 2021 IEEE International Conference on Industrial

Technology, Valencia, Spain. (ICIT 2021).

Hartford, J. S., Wright, J. R., and Leyton-Brown, K. (2016).

Deep learning for predicting human strategic behav-

ior. In Lee, D. D., Sugiyama, M., Luxburg, U. V.,

Guyon, I., and Garnett, R., editors, Advances in Neu-

ral Information Processing Systems 29, pages 2424–

2432. Curran Associates, Inc.

Hornauer, S., Yellapragada, B., Ranjbar, A., and Yu, S.

(2019). Driving scene retrieval by example from

large-scale data. In Proceedings of the IEEE/CVF

Conference on Computer Vision and Pattern Recog-

nition (CVPR) Workshops.

Kharrazi, S., Frisk, E., and Nielsen, L. (2019). Driving

behavior categorization and models for generation of

mission-based driving cycles. IEEE Intelligent Trans-

portation Systems Conference, pages 1349–1354.

Lin, H. Y., Dai, J. M., Wu, L. T., and Chen, L. Q. (2020).

A vision based driver assistance system with for-

ward collision and overtaking detection. Sensors,

20(18):100–109.

Liu, H., Taniguchi, T., Takano, T., Tanaka, Y., Takenaka,

K., and Bando, T. (2014). Visualization of driving

behavior using deep sparse autoencoder. pages 1427–

1434.

Liu, H., Taniguchi, T., Tanaka, Y., Takenaka, K., and Bando,

T. (2017). Visualization of driving behavior based

on hidden feature extraction by using deep learning.

IEEE Transactions on Intelligent Transportation Sys-

tems, 18(9):2477–2489.

Ma, W., Liu, Y., jing Zhao, and Wu, N. (2017). Increasing

the capacity of signalized intersections with left-turn

waiting areas. Transportation Research Part A: Policy

and Practice, 105:181–196.

Malekian, R., Moloisane, N. R., Nair, L., Maharaj, B., and

Chude-Okonkwo, U. A. (2014). Design and imple-

mentation of a wireless obd ii ﬂeet management sys-

tem. IEEE Sensors Journal, 13:1154–1164.

Naito, M., Miyajima, C., Nishino, T., Kitaoka, N., and

Takeda, K. (2010). A browsing and retrieval system

for driving data. IEEE Intelligent Vehicles Symposium,

pages 1159–1165.

Pearson, K. (1900). X. on the criterion that a given sys-

tem of deviations from the probable in the case of a

correlated system of variables is such that it can be

reasonably supposed to have arisen from random sam-

pling. The London, Edinburgh and Dublin Philosophi-

cal Magazine and Journal of Science, pages 157–175.

Romera, E., Bergasa, L. M., and Arroyo, R. (2016). Need

data for driver behaviour analysis? presenting the pub-

lic uah-driveset. International Conference on Intelli-

gent Transportation Systems (ITSC), pages 387–392.

Schorr, J., Hamdar, S. H., and Silverstein, C. (2016). Mea-

suring the safety impact of road infrastructure systems

on driver behavior: Vehicle instrumentation and ex-

ploratory analysis. Journal of Intelligent Transporta-

tion Systems.

Tay, Richard, Rifaat, Shakil, Chin, and Hoong (2008). A lo-

gistic model of the effects of roadway, environmental,

vehicle, crash and driver characteristics on hit-and-run

crashes. Accident; analysis and prevention, 40:1330–

Thorndike, R. L. (1953). Who belongs in the family? Psy-

chometrika, 18(4):267–276.

Wang, Y., Qian, C., Liu, D., and Hua, J. (2019). Research on

pedestrian trafﬁc safety improvement methods at typ-

ical intersection. 2019 4th International Conference

on Electromechanical Control Technology and Trans-

portation (ICECTT), pages 190–193.

Wong, C. K. (2019). Designs for safer signal-controlled

intersections by statistical analysis of accident data at

accident blacksites. IEEE Access, 7:111302–111314.

Wu, Z., Xiong, Y., Yu, S. X., and Lin, D. (2018). Unsu-

pervised feature learning via non-parametric instance

discrimination. 2018 IEEE/CVF Conference on Com-

puter Vision and Pattern Recognition, pages 3733–

3742.

Yeh, T. W., Lin, S. Y., Lin, H. Y., Chan, S. W., Lin, C. T.,

and Lin, Y. Y. (2019). Trafﬁc light detection using

convolutional neural networks and lidar data. In 2019

International Symposium on Intelligent Signal Pro-

cessing and Communication Systems (ISPACS), pages

1–2.

Zhirong, W., A, E. A., and Stella, Y. (2018). Improving

generalization via scalable neighborhood component

analysis. European Conference on Computer Vision

(ECCV) 2018.

Driving Behavior Analysis and Trafﬁc Improvement using Onboard Sensor Data and Geographic Information

291