Modeling dangerous driving events based on in-vehicle data using Random Forest and Recurrent Neural Network

Abstract

Modern vehicles produce big data with a wide variety of formats due to missing open standards. Thus, abstractions of such data in the form of descriptive labels are desired to facilitate the development of applications in the automotive domain. We propose an approach to reduce vehicle sensor data into semantic outcomes of dangerous driving events based on aggressive maneuvers. The supervised time-series classification is implemented with Random Forest and Recurrent Neural Network separately. Our approach works with signals of a real vehicle obtained through a back-end solution, with the challenge of low and variable sampling rates. We introduce the idea of having a dangerous driving classifier as the first discriminant of relevant instances for further enrichment (e.g., type of maneuver). Additionally, we suggest a method to increment the number of driving samples for training machine learning models by weighting the window instances based on the portion of the labeled event they include. We show that a dangerous driving classifier can be used as a first discriminant to enable data integration and that transitions in driving events are relevant to consider when the dataset is limited, and sensor data has a low and unreliable frequency.

Full text

Modeling dangerous driving events based on in-vehicle data using
Random Forest and Recurrent Neural Network
Daniel Alvarez-Coello∗,†, Benjamin Klotz∗,‡, Daniel Wilms∗,
Sofien Fejji∗, Jorge Marx G´
omez†, and Rapha¨
el Troncy‡
Abstract— Modern vehicles produce big data with a wide
variety of formats due to missing open standards. Thus,
abstractions of such data in the form of descriptive labels
are desired to facilitate the development of applications in
the automotive domain. We propose an approach to reduce
vehicle sensor data into semantic outcomes of dangerous driving
events based on aggressive maneuvers. The supervised time-
series classification is implemented with Random Forest and
Recurrent Neural Network separately. Our approach works
with signals of a real vehicle obtained through a back-end
solution, with the challenge of low and variable sampling
rates. We introduce the idea of having a dangerous driving
classifier as the first discriminant of relevant instances for
further enrichment (e.g., type of maneuver). Additionally, we
suggest a method to increment the number of driving samples
for training machine learning models by weighting the window
instances based on the portion of the labeled event they include.
We show that a dangerous driving classifier can be used as a
first discriminant to enable data integration and that transitions
in driving events are relevant to consider when the dataset is
limited, and sensor data has a low and unreliable frequency.
I. INTRODUCTION
Smart devices rely on high-quality data sources which can
be located and accessed remotely. Thanks to the increasing
number of connected devices, new applications can combine
multiple domains. The automotive industry also follows this
trend with its connected vehicles [1]. Nevertheless, there is
still the need for international open standards and protocols
to enable uniform data interaction [2]. Initiatives, such as
the Data-Centric Manifesto1, show the interest in data-driven
solutions regardless of the domain. Apart from the fact that
it would be too costly to share all raw vehicle sensor data
to the cloud, an application developer would have to adapt
their application throughout the variety of many models and
brands. Also, vehicles produce big time-series data [3] which
increases the complexity of data processing pipelines handled
differently by each application.
One could extract and communicate the meaning of ve-
hicle data instead of sensor values to eliminate the need for
in-depth domain knowledge for making new applications.
This abstraction process corresponds to the transformation
of data into information and is part of a Data, Information,
Knowledge, Wisdom (DIKW) hierarchy which describes the
building blocks for reasoning [4]. Our goal is to simplify
∗BMW Research, New Technologies, Innovation. Garching, Germany.
†Department of Computer Science. University of Oldenburg. Oldenburg,
Germany. daniel.alvarez@uni-oldenburg.de
‡Department of Data Science. EURECOM. Sophia Antipolis, France.
1http://datacentricmanifesto.org
vehicle data into information that describes how the driver
and vehicle behave. Thus, we focus on modeling dangerous
driving events using machine learning to classify past time
windows. In this work, we assume that just dangerous driving
events are relevant to consider for further enrichment.
This paper is organized as follows: we discuss current ap-
proaches and models related to the generation of information
on driver behavior in section II. Our approach is discussed
in section III, followed by the implementation details in IV.
The evaluation results of the classifiers and our experiments
to test them are presented in section V. We conclude with
the principal findings and possible future directions in section
VI.
II. RELATED WORK
There are different approaches to Driver Behavior Mod-
eling (DBM). Some authors study driving patterns based
on the driver’s comfort [5], or physiological signals [6];
whereas the majority focus on cameras, in-vehicle signals,
and smartphones [7]. We consider only existing vehicle sig-
nals that could be adapted in the future to possible standards
in development such as VSS2or ontological models like
VSSo3or the driving context ontology [9].
Most of the related work focuses on individual actions
of driving. The combination of such actions defines relevant
behavioral domains such as drowsiness [10]–[12], distrac-
tion [13], [14], and aggressiveness [15]–[20] (see figure 1).
Fig. 1. Main domains and actions in Driver Behavior Modeling
A. Using Cameras
In recent years, computer vision applications using deep
neural networks have shown important advances in image
2Vehicle Signal Specification https://w3.org/auto/wg/wiki/
Vehicle_Signal_Specification_(VSS)/Vehicle_Data_
Spec
3Vehicle Signal Ontology [8]
2019 IEEE Intelligent Vehicles Symposium (IV)
Paris, France. June 9-12, 2019
978-1-7281-0559-8/19/$31.00 ©2019 IEEE 165

and video processing. In the driving context, applications
can recognize different driver actions (e.g., driver’s gaze
and head position[13], [14], eye blinking[21], yawning[10],
emotions [22], etc.) and surrounded elements (e.g., traffic
signs, pedestrians, road lanes, other vehicles, etc. [23], [24]).
Although they are not under the scope of our study, such
information could be used for further data enrichment.
B. Using Vehicle Signals
The access to vehicle signals is usually restricted and
requires specific setups and dedicated hardware such as
an On-Board Diagnostics (OBD) device. Depending on the
signals of interest, an alternative is to use a vehicle simulator.
Some authors use simulated data to detect aggressive driving
events. [15] applies SVMs and K-means clustering. Simi-
larly, [16] includes more signals and uses a semi-supervised
learning approach.
There are also efforts to characterize the driver’s profile.
[25] uses signals to identify the driving style after recogniz-
ing the type of maneuver. For this purpose, logical conditions
are applied to the sensor values to determine the status of the
vehicle. However, their dataset is not available. Martinez et
al. [26] present an approach for identifying the driver, too.
While it provides an excellent foundation to learn how to
differentiate the drivers, it does not specify the behavioral
patterns. Burton et al. use the Euclidean distance traveled
and the average speed of the vehicle to discriminate driving
styles. Driver profiling does not have the granularity we look
for, because it focuses on the behavior over time and not on
single events.
C. Using External Sensors
Since accessing vehicle data can be challenging, some
researchers opt for low-cost external devices as an alternative
(e.g., smartphones, Raspberry Pi, etc.). Such devices are
equipped with inertial sensors like accelerometers, gyro-
scope, magnetometer, camera, and others.
[11] explains the implementation of a mobile applica-
tion that detects drowsiness and aggressiveness. It rates the
driver’s behavior and provides life feedback about the driving
patterns. Drowsiness is given by lane drifting and weaving
events using computer vision, where the tracking of lane
marks of the road determine how centered the vehicle is.
On the other hand, aggressiveness is inferred purely by the
accelerometers. The number of critical events defines the
level of distraction that influences the driver. The system
works only for speeds higher than 50 km/h.
[17], [18] use a smartphone as the sensing and processing
device to classify aggressive events. They use an end-point
detection algorithm, as well as Dynamic Time Warping
which is computationally expensive. A drawback is that the
entire event needs to happen before the system can process
it. On the other hand, [27] classifies the trajectory with the
aid of a highly accurate GPS data logger as either smooth
or aggressive. It uses a mathematical model, and its solution
is not suitable for a real-time application.
Another approach [28] aims to detect dangerous driving
based on four actions: abnormal speeding, steering, weaving,
and using the phone while driving. Nevertheless, no data is
collected, and the decisions rely on experimentally prede-
fined thresholds. Similarly, [29] uses also thresholds together
with an end-point algorithm to detect driving events, obtain
statistical features, and classify them using a neural network.
One way to achieve the granularity we wanted is by
classifying the maneuvers that the driver performs. The
work by Junior et al. [19] explores this topic as well as
the application of machine learning techniques to classify
driving events based on aggressiveness. Their dataset is
publicly available which facilitates the analysis. They use
a smartphone to collect 3-axis signals from accelerometer,
gyroscope, and magnetometer. This approach was used as
the primary reference for our study. With the dataset of [19],
Carvalho et al. [20] investigate the use of Recurrent Neural
Network (RNN) to classify maneuvers.
III. APPROACH AND CLASSIFICATION ASPECTS
We considered the work from Junior et al. [19] as the basis
for our approach. After replication of their grid-searches
using their public Driving Behavior Dataset4, we corrob-
orate their conclusions that Random Forest (RF) outper-
forms Support Vector Machine, Neural Network, and Multi-
Layer Perceptron in the multi-class maneuver classification
task. Therefore, we select RF as the first technique to
test. Since we are dealing with sequences, RNN are also
considered [20].
A. Base Classifier
In contrast to [19], we propose to split the multi-class
classification problem into two parts as shown in figure 2.
A binary classifier of dangerous driving will tell us what
time-window instances are relevant for further processing.
Then other criteria (e.g., type of maneuver) can be used to
enrich the outcomes of the base classifier. In this way, results
of different applications could also be added (e.g., driver’s
emotion of the last seconds, drowsiness percentage, gaze,
etc.).
Fig. 2. A base classifier detects a dangerous situation by using relevant
vehicle signal data. More specific classifiers can enrich the outcome
B. Feature Selection and Extraction
Based on [26], [30], we selected the subset of 12 signals
shown in the table I. We have two types of variables: con-
tinuous and categorical. For continuous signals, we extract
statistical features (i.e., mean, median, standard deviation,
4https://github.com/jair-jr/driverBehaviorDataset
166

and trend [19]). For categorical signals we take only the
median value. All 12 signals were used in RF. For RNN,
we did not use 3 of the signals because they were ranked as
less important in the analysis with Random Forest: displayed
speed, gear, and brake DSC safe.
Continuous signals Categorical signals
Lateral acceleration Acceleration efficiency
Longitudinal acceleration Gear *
Accelerator pedal position Brake pressed
Actual speed Brake Dynamic Stability Control (DSC) state *
Speed displayed *
Engine consumption
Engine RPM speed
Engine torque
TABLE I
SEL ECTE D SIGNA LS. THOSE MARKED WITH “*” W ERE N OT
CONSIDERED FOR RNN
C. Instance Relevance
One driving event E, in our case a maneuver, is composed
of a sequence of measurements with a duration of esize.
To classify the events, the size of the time window wsize
should generalize for all the maneuvers of interest. The ones
we want to classify are in the range of a few seconds. Thus,
we tried out window sizes between 1 and 10 seconds.
Nevertheless, the low and irregular sampling frequency we
get at the back-end added complexity to the classification
of short driving events. Especially because sometimes time
windows do not contain enough values to constitute a sample
suitable for training. To overcome the limitation, we propose
to consider the transitions between driving events as valid
instances for training.
When Whops over time, it will not always be covering
the whole driving event (i.e., Wpartially overlapping E). A
window instance is when Wis at a specific time step. We
only care about instances when the window is overlapping
the occurrence of E. With that said, the total number of
instances of one driving event instancestotal, and the
instance index iare given by:
instancestotal =wsize +esize −1(1)
i∈ {1,2,3, ..., (wsize +esize −1)}(2)
Since one maneuver has several instances, such instance
have different relevance. To determine the importance of the
instances, we introduce a method to calculate their relevance
as a number between 0 (not important) and 1 (most relevant).
It considers the following aspects:
•When wsize =esize , there exists just one instance with
a relevance of 1.
•When wsize > esize , relevance is 1 for all instances in
which all the frames of Eare covered by W.
•When wsize < esize , relevance is 1 for instances in
which Wis inside E.
relevance =n
wsize
·n
esize
·k=n2
wsize ·esize
·k(3)
n=


i+ 1,if (i<wsize )and (i<esize )
i−1,if (i>wsize )and (i>esize )
1,otherwise
(4)
k=


wsize
esize ,if (wsize > esiz e)
esize
wsize ,if (wsize < esiz e)
1,if (wsize =esize )
(5)
D. Random Forest Parameters
In addition to [19], we add the window size and the
instance relevance as our custom parameters. We do a grid-
search for RF to find the best parameters from the table II.
Custom parameters
Window size [frames] {2, 3, 4, ..., 10}
Minimum instance relevance {0.1, 0.2, ..., 1.0}
Random Forest
Number of estimators {10, 11, 12, ..., 25}
Maximum features {10, 15, “log2”}
Maximum depth {5, 10, 15}
TABLE II
PARA ME TE RS T ES TE D IN T HE G RI D -SEARCH FOR RF
E. Recurrent Neural Network Parameters
For RNN, we trained different combinations of parameters
based on [20] (see table III). We used a window size of
10 frames and an instance relevance of 0.7. The number of
epochs was 500 with an early stopping patience of 50 epochs.
The optimizer was “RMSprop” and the learning rate 0.001.
Recurrent Neural Network
Number of hidden layers {1, 2}
Number of recurrent units in the hidden layer {10, 15, 16, 32, 64, 128}
Recurrent unit type {LSTM, GRU}
Dropout {0.1, 0.2}
Recurrent dropout {0.1, 0.2}
TABLE III
PARAMETERS TESTED ON THE RNN IMPLEMENTATION
IV. IMPLEMENTATION
A. Dataset
We collected and labeled vehicle data of two licensed
drivers. The maneuvers were the same as in [19] (i.e.,
aggressive and normal turns, lane changes, accelerations, and
braking). The signals come from the CAN which is accessed
via a dedicated back-end architecture that is developed within
BMW Research. The collection of data is only used in
research for tests like the one conducted in this study.
Once the data was collected, we downsampled the series
to half-second periods by assigning the aggregated values to
167

the starting point of the current time window. This frequency
was determined based on the lowest rate among the selected
signals.
B. Considerations for Training
•We use the Area Under the ROC curve (AUC) [31] as
the evaluation metric for the trained models since it is a
trade-off between False Positive Rate and True Positive
Rate by considering all the possible thresholds for the
classification. AUC is a better metric for classification
problems that have an imbalanced number of samples.
•We joined left and right lane-changes to deal with low
sampling frequency issues because those maneuvers
are the shortest in duration. Lateral acceleration was
inverted to double the number of samples in this class.
•We used binary cross-entropy as the loss function
for driving mode classification and categorical cross-
entropy for the maneuver classification.
•To use RNNs, we first normalized our data according
to the minimum and maximum possible values of the
signals. The input layer is feed with sequences of 10
measurements from 9 signals.
V. RESULTS
In this section, we present the results of the best classi-
fiers found for our specific use-case and the corresponding
experiments that were conducted.
A. Classifiers Evaluation
The grid-search on RF showed us that models which use
instance relevance were ranked among the best combinations.
The parameters corresponding to the best RF found are
presented in table IV.
Parameter \Classifier Base Maneuver
Window size [frames] 10 10
Minimum instance relevance 0.9 0.8
Number of estimators 15 24
Maximum features 5 “log2”
Maximum depth 10 15
TABLE IV
PARA ME TE RS O F TH E BE ST R F FOUND
Regarding RNN, one hidden layer with 64 recurrent units
showed a better score for both classification tasks. The output
layer contains 2 and 5 units for driving mode and maneuvers
respectively. LSTM cells did slightly better than GRU for
most of the tested combinations.
For the base classifier, both RF and RNN classified the
instances of the test set correctly. While for the maneuver
classification, the corresponding confusion matrices and the
ROC curves show a few miss classifications of turns and
lane changes (see figure 3). Nevertheless, the consideration of
lane-changes to both sides as just one class (doubled samples
by inverting the sign of the lateral acceleration) improved
the performance significantly compared to the first attempt
when we tried to classify lane change to the right and left
separately.
(a) Normalized confusion matrix
(b) Multi-class ROC curve
Fig. 3. Evaluation of dangerous maneuver classifier using RNN
B. Test Experiments
We tested the best found models with 10 unseen trajec-
tories. For this purpose, we used two different routes of a
track where each trajectory corresponded to one lap. The
test drivers were given specific instructions (see table V) on
how to drive before each lap. The instruction of 3 driving
styles refers to 3 laps on a given route, where each lap had
a different style (i.e., normal, moderate, aggressive).
For every time step, the base classifier will predict the
class of the previous 10 frames, the overall danger score is
168

Route Driver Instruction
1 A 3 driving styles
1 B 3 driving styles
1 B 2 laps of free driving
2 A 3 driving styles
TABLE V
TEST TRAJECTORIES AND THEIR CORRESPONDING INSTRUCTION
calculated by dividing the total number of positive outcomes
by the total number of time steps. If we want to have more
granularity, we can calculate a moving score by considering
only a given amount of previous time steps. As we see
in figure 4, the overall danger score of the RNN base
classifier reflects the instructions given to the driver for the
3 driving styles: normal (lower-level), moderate (mid-level),
and aggressive (upper-level). Likewise, the moving score
provides us more insight into how the behavior over time
is.
(a) Overall danger score
(b) Score of the past 50 frames
Fig. 4. Danger score of Driver A on route 1 (3 driving styles) using RNN
Additionally, we collected 2 laps of free driving and
compared the model’s outcome against the perception of two
co-pilots who were inside the vehicle. The co-pilots wrote
down their perception of danger in a scale from 0 (no danger)
to 4 (most dangerous). The average danger score perceived
by the co-pilots was roughly 67%, which is not far from the
approximately 75% predicted by the models (see figure 5).
Since we know the sequences of the maneuvers performed
on the track, one can map the outcomes of an unseen
trajectory (i.e., data that is new to the trained model) to check
how consistent they are. Figure 6 shows how the classified
maneuvers of an aggressive lap match the sequences of the
track.
Fig. 5. Danger score of Driver B on route 1 (2 laps of free driving)
VI. CONCLUSION
A binary classifier of dangerous driving events based on
in-vehicle signals can simplify vehicle data and enable the
integration of other domains. Compared to state-of-the-art
methods, the proposed approach can provide similar results
with lower and variable sampling rate. The instance relevance
let us use samples in which a hopping window overlaps the
driving event partially.
The collection of dangerous driving samples is time-
consuming and requires special considerations (i.e., dedi-
cated track, qualified drivers, correct labeling, etc.). A limi-
tation of our implementation is that the labels were selected
by one person, which translates into a model that represents
the labeler’s danger perception. Hence, the assignation of
labels should be extended to more criteria to reduce potential
bias. To overcome this situation we are working on an
infrastructure to involve normal test drivers outside dedicated
tracks. Mainly, we use the base classifier model to recognize
dangerous driving events in the back-end, let the driver rate
the detected situation and reinforce the model over time with
less overhead.
Sending classified driving events over the network is more
practical than transferring raw data. Therefore, our next
steps would be to predict incoming data streams directly
in the vehicle with a reinforced model and use graph data
to prioritize data relationships for the integration with other
domains. One approach to deal with interoperability issues
across platforms could be by mapping the detected driving
events to a standardized data model, such as VSS/VSSo.
REFERENCES
[1] Cisco, “Internet of things,” 2016, accessed: 2019-01-25. [Online].
Available: www.bit.ly/2vpGRxp
[2] Postscapes, “Iot standards and protocols,” 2018, ac-
cessed: 2018-05-03. [Online]. Available: www.postscapes.com/
internet-of- things-protocols/
[3] McKinsey, “Ready for inspection – the automotive aftermarket in
2030,” McKinsey Center for Future Mobility, June 2018. [Online].
Available: www.bit.ly/2MEhXlG
[4] J. Rowley, “The wisdom hierarchy: representations of the dikw hier-
archy,” Journal of information science, vol. 33, no. 2, pp. 163–180,
2007.
[5] X. Chang, J. Rong, C. Zhou, and H. Li, “Relationship between
driver’s feeling and vehicle operating characteristics on urban road,”
in Intelligent Control and Automation (WCICA), 2016 12th World
Congress on. IEEE, 2016, pp. 3033–3037.
[6] N. Li, T. Misu, and A. Miranda, “Driver behavior event detection for
manual annotation by clustering of the driver physiological signals,”
in Intelligent Transportation Systems (ITSC), 2016 IEEE 19th Inter-
national Conference on. IEEE, 2016, pp. 2583–2588.
169

Fig. 6. Trajectory reconstruction of route 1 from the aggressive lap of Driver A using RNN
[7] N. AbuAli and H. Abou-zeid, “Driver behavior modeling: Devel-
opments and future directions,” International Journal of Vehicular
Technology, 2016.
[8] B. Klotz, R. Troncy, D. Wilms, and C. Bonnet, “VSSo - A
vehicle signal and attribute ontology,” in SSN 2018, 9th International
Semantic Sensor Networks Workshop, 9 October 2018, Monterey, CA,
USA, Monterey, UNITED STATES, 10 2018. [Online]. Available:
http://www.eurecom.fr/publication/5691
[9] ——, “A driving context ontology for making sense of cross-domain
driving data,” in BMW Summer school, Raitenhaslach, Germany, July-
August 2018.
[10] B. Reddy, Y.-H. Kim, S. Yun, C. Seo, and J. Jang, “Real-time driver
drowsiness detection for embedded system using model compression
of deep neural networks,” in Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition Workshops, 2017, pp. 121–
128.
[11] L. M. Bergasa, D. Almer´
ıa, J. Almaz´
an, J. J. Yebes, and R. Arroyo,
“Drivesafe: An app for alerting inattentive drivers and scoring driv-
ing behaviors,” in Intelligent Vehicles Symposium Proceedings, 2014
IEEE. IEEE, 2014, pp. 240–245.
[12] E. Romera, L. M. Bergasa, and R. Arroyo, “Need data for driver
behaviour analysis? presenting the public uah-driveset,” in Intelligent
Transportation Systems (ITSC), 2016 IEEE 19th International Confer-
ence on. IEEE, 2016, pp. 387–392.
[13] T. Liu, Y. Yang, G.-B. Huang, Y. K. Yeo, and Z. Lin, “Driver
distraction detection using semi-supervised machine learning,” IEEE
transactions on intelligent transportation systems, vol. 17, no. 4, pp.
1108–1120, 2016.
[14] I.-H. Choi, S. K. Hong, and Y.-G. Kim, “Real-time categorization of
driver’s gaze zone using the deep learning techniques,” in Big Data
and Smart Computing (BigComp), 2016 International Conference on.
IEEE, 2016, pp. 143–148.
[15] W. Wang and J. Xi, “A rapid pattern-recognition method for driving
styles using clustering-based support vector machines,” in American
Control Conference (ACC), 2016. IEEE, 2016, pp. 5270–5275.
[16] W. Wang, J. Xi, A. Chong, and L. Li, “Driving style classification
using a semisupervised support vector machine,” IEEE Transactions
on Human-Machine Systems, vol. 47, no. 5, pp. 650–660, 2017.
[17] D. A. Johnson and M. M. Trivedi, “Driving style recognition using
a smartphone as a sensor platform,” in Intelligent Transportation
Systems (ITSC), 2011 14th International IEEE Conference on. IEEE,
2011, pp. 1609–1615.
[18] H. Eren, S. Makinist, E. Akin, and A. Yilmaz, “Estimating driving
behavior by a smartphone,” in Intelligent Vehicles Symposium (IV),
2012 IEEE. IEEE, 2012, pp. 234–239.
[19] J. F. J´
unior, E. Carvalho, B. V. Ferreira, C. de Souza, Y. Suhara,
A. Pentland, and G. Pessin, “Driver behavior profiling: An investiga-
tion with different smartphone sensors and machine learning,” PLoS
one, vol. 12, no. 4, p. e0174959, 2017.
[20] E. Carvalho, B. V. Ferreira, J. Ferreira, C. de Souza, H. V. Carvalho,
Y. Suhara, A. S. Pentland, and G. Pessin, “Exploiting the use of
recurrent neural networks for driver behavior profiling,” in Neural
Networks (IJCNN), 2017 International Joint Conference on. IEEE,
2017, pp. 3016–3021.
[21] A. Rosebrock, “Drowsiness detection with OpenCV,” 2017, accessed:
2018-01-15. [Online]. Available: www.pyimagesearch.com/2017/05/
08/drowsiness-detection-opencv/
[22] O. Arriaga, M. Valdenegro-Toro, and P. Pl ¨
oger, “Real-time convolu-
tional neural networks for emotion and gender classification,” arXiv
preprint arXiv:1710.07557, 2017.
[23] O. Kumtepe, G. B. Akar, and E. Yuncu, “Driver aggressiveness
detection via multisensory data fusion,” EURASIP Journal on Image
and Video Processing, vol. 2016, no. 1, p. 5, 2016.
[24] S. M. Hegazy and M. N. Moustafa, “Classifying aggressive drivers for
better traffic signal control.”
[25] M. Van Ly, S. Martin, and M. M. Trivedi, “Driver classification and
driving style recognition using inertial sensors,” in Intelligent Vehicles
Symposium (IV), 2013 IEEE. IEEE, 2013, pp. 1040–1045.
[26] M. V. Mart´
ınez, I. Del Campo, J. Echanobe, and K. Basterretxea,
“Driving behavior signals and machine learning: a personalized driver
assistance system,” in Intelligent Transportation Systems (ITSC), 2015
IEEE 18th International Conference on. IEEE, 2015, pp. 2933–2940.
[27] A. B. R. Gonz´
alez, M. R. Wilby, J. J. V. D´
ıaz, and C. S. ´
Avila,
“Modeling and detecting aggressiveness from driving signals,” IEEE
Transactions on intelligent transportation systems, vol. 15, no. 4, pp.
1419–1428, 2014.
[28] F. Li, H. Zhang, H. Che, and X. Qiu, “Dangerous driving behavior
detection using smartphone sensors,” in Intelligent Transportation
Systems (ITSC), 2016 IEEE 19th International Conference on. IEEE,
2016, pp. 1902–1907.
[29] P. Brombacher, J. Masino, M. Frey, and F. Gauterin, “Driving event de-
tection and driving style classification using artificial neural networks,”
in Industrial Technology (ICIT), 2017 IEEE International Conference
on. IEEE, 2017, pp. 997–1002.
[30] K. Zfnebi, N. Souissi, and K. Tikito, “Driver behavior quantitative
models: Identification and classification of variables,” in Networks,
Computers and Communications (ISNCC), 2017 International Sympo-
sium on. IEEE, 2017, pp. 1–6.
[31] A. P. Bradley, “The use of the area under the roc curve in the evaluation
of machine learning algorithms,” Pattern recognition, vol. 30, no. 7,
pp. 1145–1159, 1997.
170