See discussions, stats, and author profiles for this publication at: [Link]

net/publication/357447981

Traffic Density Estimation using Machine Learning Methods

Article · December 2021

CITATION READS

1 1,201

3 authors, including:

Sümeyye Ünlüleblebici Murat Taşyürek

Kayseri University Kayseri University
7 PUBLICATIONS 6 CITATIONS 60 PUBLICATIONS 309 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Murat Taşyürek on 31 December 2021.

The user has requested enhancement of the downloaded file.

Journal of Artificial Intelligence and Data Science (JAIDA) Vol. 1, No. 2, pp. 136-143, 2021
e-ISSN: 2791-8335 Research Article
[Link]

Traffic Density Estimation using Machine Learning Methods

Sümeyye AYDIN 1,*, Murat TAŞYÜREK 1, Celal ÖZTÜRK 2
1*
Kayseri University, Engineering, Architecture And Design Faculty, Computer Engineering Department, Turkey
2
Erciyes University, Engineering Faculty, Computer Engineering Department, Turkey

Abstract
In cities where population density is high and transportation systems are widely used, it is necessary to manage
traffic systems more effectively not to affect the daily planned works. The Intelligent Transportation System
(AUS) is expressed as a system that provides users with better information and safer, more coordinated, and
smarter use of transportation networks with different transportation modes and traffic management. One of the
most important components of AUS models is the determination of traffic density. The traffic density of
intersections is a difficult problem as it affects other interconnected intersections and varies in time. Deep
learning method is a widely used method in traffic density estimation in recent years. In this study, the long-
term short memory network (LSTM) model, one of the deep learning methods, is proposed to estimate the
traffic density of a certain region using open data of Istanbul Metropolitan Municipality. The performance of
the proposed LSTM-based model is compared with machine learning methods such as linear regression,
decision tree, random forest, and the classical deep learning method (DL). Experimental evaluations show that
the proposed LSTM method is more successful in traffic density estimation than the compared methods.
Keywords: Deep learning, Intelligent transportation systems; Istanbul traffic density prediction; LSTM,
Machine Learning.

1. Introduction
Intelligent Transportation Systems (ITS) technologies improve transportation safety and mobility and increase
productivity by integrating advanced communication technologies into transportation infrastructure and vehicles.
One of the most important components of ITS technologies is the estimation of traffic density. Traffic density is
expressed as the traffic demand exceeding the transportation capacity. The traffic density of one intersection is a
difficult problem as it also affects other interconnected intersections and varies in time. Traffic density is a serious
problem that affects the planning of daily life and is becoming increasingly common all over the world. This paper
proposes to estimate the intersection density in a particular region by using LSTM-based deep learning method
with open data of Istanbul Metropolitan Municipality (IMM).
Deep learning methods have recently been widely used to solve many problems such as natural language
processing and real-time object detection [1-3]. There are many different deep learning models in the literature,
depending on the data set and the type of problem. Some deep learning methods are insufficient in time-dependent
processes. Recurrent Neural Networks (RNN) are successful in datasets containing time-dependent relationships
[4]. The RNN model was first introduced in the 1980s [5]. However, traditional RNN methods face the problem
of capturing long-term dependencies in the input data. A long short-term memory network (LSTM) has been
proposed to solve this problem [6]. Ma et al. have applied the LSTM model for estimating traffic speed with
remote micro software sensor data [7]. Tian et al. proposed the LSTM RNN model for traffic flow prediction [8]
and found that the LSTM RNN method outperformed most of the non-parametric models [8]. Li et al. evaluated
the performance of the LSTM and GRU model for traffic flow prediction [9]. A better performance can be
obtained with the LSTM model in traffic flow forecasting, as LSTM can automatically calculate optimal time
delays and capture features of longer time interval time series [8].
In this study, it is proposed to estimate traffic density using LSTM-based deep learning method using IMM
open transportation data. Istanbul is the city with the largest population density and the most significant traffic
problem in Turkey [10,11]. The proposed method is aimed to predict the traffic density of certain intersections
during the time periods when the traffic density is high, such as commuting and rush hours. The performance of
the proposed LSTM-based model is compared with machine learning methods such as linear regression, decision
tree, random forest, and the classical deep learning method (DL) on actual observed data.

*Corresponding author e-mail address: sumeyyeaydin@[Link]

136
Received: 29/11/2021; Accepted: 21/12/2021
 AYDIN et al. / JAIDA vol (2021) 136-143

2. Methods

2.1. Linear regression method

Linear regression aims to establish a linear relationship between one or more independent variables and a
dependent variable or a numerical result. This method estimates the value according to the independent variables.
In simple linear regression, the target value can be estimated with one independent variable, while in multiple
regression, the target is estimated by two or more independent variables. A multiple linear regression model was
used because the number of independent variables in the traffic dataset was more than one.

2.2. Decision tree

Decision trees are a hierarchical machine learning method consisting of nodes, branches, and leaves. The
output of the decision tree is in the form of a tree or rules, so it is easy to understand. The tree expands gradually
according to the questions asked to the root node, and the growth is completed when the last leaves are formed. It
is divided into regression trees and classification trees. If the dependent variable is categorical, classification trees
are used, and if the dependent variable is continuous, regression trees are used. In this article, regression trees
were used because it works with constantly changing data.

2.3 Random forest

Random forest algorithm is a machine learning method used to solve regression and classification
problems[12]. The algorithm is an ensemble learning algorithm consisting of the output of multiple decision trees.
Each node branches by choosing the best among the randomly selected variables in the node [13]. It is not pruned
as in decision trees, so it is fast. In this study, a forest was created using 30 decision trees.

2.4. Deep learning

Deep learning is a subset of machine learning. It has many neural nodes. An artificial neural network is formed
by the combination of nerve cells. A neural network is called ‘deep’ when it has more than one hidden
layer[14].This design was inspired by the biological structure of the human brain. So in deep learning, there can
be many layers between the input and output layers. These layers are called hidden layers. In the model consisting
of consecutive layers, the previous layer's output is the input of the next layer[15]. Deep learning has been a
widely used prediction method in recent years.

2.5. Long short-term memory (LSTM)

The LSTM deep learning algorithm is known as a recurrent neural network introduced by Hochreiter and
Schmidhuber in 1997 to eliminate the disadvantages of the RNN architecture [16]. The most important feature of
the LSTM network is that it also evaluates the background information about the network. The basic LSTM
diagram is given in Figure 1 [17, 18].

Figure 1. Basic LSTM block diagram.

LSTM networks address the problem of vanishing gradients of the RNN model by dividing into three inner
cell gates, creating pseudo-memory cells to store information in a long-range context [19]. A typical LSTM neural
network cell is basically configured with four gates, namely the input gate, the input modulation gate, the forget
gate, and the output gate: The cell is controlled by gates and can keep the value or reset the value according to the
state of the gate. In particular, three gates are used to control whether the current cell value will be forgotten
137
 AYDIN et al. / JAIDA vol (2021) 136-143

(forget gate 𝑓𝑡 ), whether it will read its input (input gate 𝑖𝑡 ), and output (output gate 𝑜𝑡 ) the new cell value; In
addition, there is an input modulation gate called c̅ 𝑡 . The LSTM gates and their operations are presented in Eqs.
(1)-(6).
𝑓𝑡 = 𝜎(𝑊𝑥𝑓 𝑥𝑡 + 𝑊ℎ𝑓 ℎ𝑡−1 + 𝑏𝑓 ) (1)
i𝑡 = 𝜎(𝑊𝑥𝑖 𝑥𝑡 + 𝑊ℎ𝑖 ℎ𝑡−1 + 𝑏𝑖 ) (2)
c̅ 𝑡 = 𝑡𝑎𝑛ℎ(𝑊𝑥𝑐 𝑥𝑡 + 𝑊𝑐ℎ ℎ𝑡−1 + 𝑏𝑐 ) (3)
c𝑡 = 𝑓𝑡 𝑐𝑡−1 + 𝑖𝑡 c̅ 𝑡 ) (4)
o𝑡 = 𝜎(𝑊𝑥𝑜 𝑥𝑡 + 𝑊ℎ𝑜 ℎ𝑡−1 + 𝑏𝑜 ) (5)
h𝑡 = o𝑡 ∗ tanh(c𝑡 ) (6)
The gateway receives a new entry point from the outside and processes the newly incoming data. The memory
cell input gate receives input from the output of the LSTM neural network cell in the last iteration. The forget gate
decides when to forget the output results and thus selects the optimal time delay for the input sequence. The output
port receives all the calculated results and generates the output for the LSTM neural network cell [6]. LSTM
models redesign computing nodes based on the structure of the RNN network.

3. Experimental Evaluation

3.1. Dataset
In this study, traffic density data of August 2020 belonging to IMM were used [20]. IMM open dataset includes
time information, locations of intersections, number of vehicles based on intersections, average speed, maximum
speed and minimum speed of these vehicles. The dataset used in this study contains 1.048.575 rows of records.
The spatial locations of the intersections used in this study are shown in Figure 2 and the summary information
of the dataset is presented in Table 1.

Figure 2. Location of target stations in the used dataset

When the IMM open dataset, the general information of which is presented in Table 1, is examined in detail,
an increase was observed in the number of vehicles, especially during the commuting and exiting hours, while a
decrease was observed during the night hours. This paper shows that the number of vehicles in traffic changes
over time. Since the traffic density changes according to time, the data were sorted on the basis of date and time

138
 AYDIN et al. / JAIDA vol (2021) 136-143

before the training process. In addition, the clock data is added to the dataset as an input. In determining the
number of vehicles in traffic, it is not sufficient to know only the vehicle number data of a certain time before the
relevant station. An intersection is also affected by the density of intersections close to it. An intersection usually
consists of 4 branches, so it is important to know the four intersections close to each intersection. Therefore,
giving the number of vehicles in four intersections close to the target intersection where the traffic density will be
estimated as an input to the LSTM model will strengthen the prediction ability. For this reason, the 4 intersections
that are spatially closest to each intersection are determined by using the Euclidean distance as the spatial distance.
After determining the four intersections, which are the closest neighbours of each intersection, the number of
vehicles in the same date and time zone in these neighbouring intersections is used as input data for the intersection
whose density will be estimated.

Table 1. Data Dictionary.

Column Type Tag Definition
DATE_TIME text Date It is the field that contains the date and time information. The data
format is YYYY-MM-DD HH24:MI:SS. * Date breakdown is hourly.
LONGITUDE text Longitude It is the field that contains the longitude information.
LATITUDE text Latitude It is the field that contains the latitude information.
GEOHASH text geohash Geohash Value of Latitudes and Longitudes
MINIMUM_SPEED text Minimum Speed Minimum speed (km/h) in the geohash area at the relevant hour.
MAXIMUM_SPEED text Maximum speed Maximum speed (km/h) in the geohash area at the relevant hour.
AVERAGE_SPEED text Average Speed Average speed (km/h) in the geohash area at the relevant hour.
NUMBER_OF_VEHICLES text Number of Number of different vehicles in the geohash area at the relevant hour.
Individual Vehicles

3.2 Model Settings

By working on the number of layers in the LSTM model, the most suitable LSTM input layer was determined
to be one output layer and 20 neurons. The structure of the LSTM model used in the study is given in Figure 3.

Figure 3. LSTM model structure in the study.

In the study with the LSTM method, the Adam algorithm was used as the optimization method. Adam is an
optimization algorithm that can be used instead of the classical stochastic gradient descent procedure to update
iterative Evaluation Metrics mesh weights based on training data.

3.3 Evaluation Metrics

In this study, MAE and RMSE measurements were used as evaluation metrics for performance comparison.
The study was carried out by normalizing the data, in order not to make an incorrect comparison, the data was
converted to real values by reverse scaling and evaluated.
The MAE is the mean of the absolute error between the predicted value and the true value. It ranges from zero
to infinity, it is expected to give a low result. It is a linear value. The MAE formula is shown in Eq. (7). In the
formula, 𝑝𝑖 represents the predicted value and 𝑜𝑖 original value.

1
𝑀𝐴𝐸 = 𝑛 ∑𝑛𝑖=1|𝑜𝑖 − 𝑝𝑖 | (7)

139
 AYDIN et al. / JAIDA vol (2021) 136-143

The RMSE is the root mean square error between the predicted value and the true value. It ranges from zero
to infinity, it is expected to give a low result. The RMSE formula is shown in Eq. (8). In the formula, 𝑝𝑖 represents
the predicted value and 𝑜𝑖 original value.
1
𝑅𝑀𝑆𝐸 = √𝑛 ∑𝑛𝑖=1(𝑜𝑖 − 𝑝𝑖 )2 (8)

3. Experiment Results
Experimental evaluations of linear regression, decision trees, random forest, classical deep learning, and
LSTM methods were made using real data from IMM. RMSE and MAE values were used as evaluation criteria
to examine the performance of the methods. The station information and the values of the methods that were
experimentally evaluated are presented in Table 2 and Table 3.
When the experimental evaluation results (Table 2) created according to the MAE error values are examined,
the LSTM method generally gave lower error results than the other compared methods.

Table 2. MAE Values of the Methods.

İstasyon Geohash Decision Tree Deep Learning Linear Regression LSTM Random Forest
sxk3q9 5,325 3,799 4,237 3,816 3,773
sxk3r9 25,678 29,115 19,625 15,152 16,536
sxk3rf 12,864 12,387 10,574 9,96 9,878
sxk3w2 5,019 4,883 4,315 4,282 4,431
sxk3w6 39,309 32,604 25,265 23,649 24,429
sxk90y 17,365 16,137 14,353 13,251 13,531
sxk91n 5,837 5,311 5,112 4,745 4,702
sxk91u 9,289 7,3 7,637 7,111 7,493
sxk91w 8,821 6,984 6,386 5,317 6,939
sxk92b 10,835 8,438 9,205 8,753 8,764
sxk92c 7,1 6,053 5,633 5,582 5,998
sxk92e 7,117 4,547 4,592 4,606 4,757
sxk92f 8,301 6,155 6,166 5,71 6,012
sxk92q 8,926 8,218 8,231 7,65 7,739
sxk93s 13,367 13,304 11,497 8,803 11,49
sxk966 8,387 5,696 5,752 5,532 5,948
sxk96e 6,952 5,193 5,444 5,118 5,52
sxk96h 24,372 24,022 19,673 17,577 21,22
sxk99b 9,475 7,279 8,53 7,291 7,449
sxk99y 10,687 8,745 7,205 6,225 7,664

Table 3 shows statistics on the RMSE values of the methods.

Table 3. RMSE Values of the Methods.

Station Geohash Decision
Deep Learning Linear Regression LSTM Random Forest
Tree
sxk3q9 7,429 4,748 5,135 4,797 4,778
sxk3r9 35,96 37,584 24,058 20,58 20,449
sxk3rf 17,084 16,447 13,687 13,353 13,093
sxk3w2 6,405 6,165 5,317 5,405 5,471
sxk3w6 54,811 41,776 32,909 31,34 30,639
sxk90y 21,559 20,193 18,814 16,812 17,948
sxk91n 7,658 5,763 6,465 6,074 5,978
sxk91u 11,643 8,658 9,506 8,872 9,303
sxk91w 10,953 7,07 7,897 7,164 8,358
sxk92b 13,44 10,741 11,763 11,321 11,339
sxk92c 9,411 7,757 7,387 7,358 7,92

140
 AYDIN et al. / JAIDA vol (2021) 136-143

sxk92e 8,838 6,276 5,948 6,16 5,954

sxk92f 10,008 7,234 7,844 7,468 7,643
sxk92q 12,272 9,66 10,427 9,777 10,162
sxk93s 17,194 16,429 14,337 13,007 14,445
sxk966 11,249 6,577 7,471 7,261 7,95
sxk96e 8,927 7,934 7,067 6,706 7,309
sxk96h 30,717 31,072 23,533 21,894 26,93
sxk99b 12,262 9,885 10,876 9,834 9,803
sxk99y 14,047 14,418 8,935 8,205 9,734

When the result values were examined (Table 2 and Table 3), the LSTM method made predictions with lower
error rates compared to other methods. In addition, comparison of MAE values with statistical data is presented
in Table 4. Standard deviation value of the LSTM method gave better results compared to other methods (Table
4). This shows that the LSTM method works more stable.

Table 4. Comparison of MAE error values.

Method Total Avg Best Worst Std.
Decision Tree 245,03 12,25 5,02 39,31 8,53
Deep Leaning 216,17 10,81 3,80 32,60 8,38
Linear Regression 189,43 9,47 4,24 25,27 5,89
LSTM 170,13 8,51 3,82 23,65 5,16
Random Forest 184,27 9,21 3,77 24,43 5,65

In Figure 4, the success of LSTM is seen with the low total RMSE value in the traffic density of the stations.

Figure 4. Total RMSE values of the method.

Figure 5 shows the actual number of vehicles and the number of vehicles estimated by decision trees, deep
learning, linear regression, LSTM and random forest methods in the sxk99y geohash station. Comparison of
methods in the sxk99y geohash station is given in figure [Link] graph with the highest overlap between the
predicted value and the actual value belongs to the study using the LSTM method.

141
 AYDIN et al. / JAIDA vol (2021) 136-143

Figure 5. Comparison of methods test and prediction data for station sxk3r9.

Conclusion
The traffic problem takes a large part of people's time in last years. In addition, it is very important to manage
traffic systems in places with high population density. Intelligent Transportation Systems (ITS) are used to
manage traffic systems to overcome the traffic density problem. One of the most important components of AUS
models is that traffic density can be determinated in advance. For this purpose, a model has been proposed to
predict the number of vehicles in the traffic by working on the data of the city of Istanbul, which is the most
crowded and has the most traffic in Turkey.
LSTM model, which is a deep learning approach, is proposed for traffic density estimation. Since LSTM is
able to automatically calculate optimal time delays and capture the characteristics of longer time-lapse time series,
it has achieved higher performance in traffic flow forecasting. Experimental evaluations have shown that LSTM-

142
 AYDIN et al. / JAIDA vol (2021) 136-143

based methods are more successful than linear regression, decision trees and random forest, and the classical deep
learning methods.

Declaration of Interest
The authors declare that there is no conflict of interest.

Acknowledgements
An earlier version of this paper was presented at the ICADA 2021 Conference and was published in its Proceedings (Title
of the conference paper: “Uzun Süreli Kısa Bellek Ağı Yöntemi ile Trafik Yoğunluk Tahmini”).
We would like to thank IMM for sharing the hourly traffic density data used in this study.

References
[1] D. Ravì, C. Wong, F. Deligianni, M. Berthelot; J. Andreu-Perez, B. Lo and G. Yanget, "Deep Learning for Health Informatics," in
IEEE Journal of Biomedical and Health Informatics, vol. 21, no. 1, pp. 4-21, doi: 10.1109/JBHI.2016.2636665, Jan. 2017
[2] T. Young, D. Hazarika, S. Poria and E. Cambria, "Recent Trends in Deep Learning Based Natural Language Processing [Review
Article]," in IEEE Computational Intelligence Magazine, vol. 13, no. 3, pp. 55-75, doi: 10.1109/MCI.2018.2840738, Aug. 2018.
[3] A. Şeker, B. Diri ve H. H. Balık, "Derin Öğrenme Yöntemleri Ve Uygulamaları Hakkında Bir İnceleme", Gazi Mühendislik
Bilimleri Dergisi (GMBD), c. 3, sayı. 3, ss. 47-64, Aralık 2017
[4] A. J. P. Samarawickrama, T. G. I. Fernando, "A recurrent neural network approach in predicting daily stock prices an application
to the Sri Lankan stock market," 2017 IEEE International Conference on Industrial and Information Systems (ICIIS), pp. 1-6, doi:
10.1109/ICIINFS.2017.8300345, 2017
[5] N. Buduma and N. Locascio, Fundamentals of Deep Learning. Designing Next-Generation Machine Intelligence Algorithms,
O'Reilly Media, 172-217, 2017.
[6] S. Hochreiter and J. Schmidhuber, "Long Short-Term Memory," in Neural Computation, vol. 9, no. 8, pp. 1735-1780, 15 Nov.
1997, doi: 10.1162/neco.1997.9.8.1735.
[7] X. Ma, Z. Tao, Y. Wang, H. Yu, and Y. Wang, “Long short-term memory neural network for traffic speed prediction using remote
microwave sensor data” Transportation Research Part C: Emerging Technologies, 54, pp. 187–197, 2015.
[8] Y. X. Tian and P. Li, "Predicting Short-term Traffic Flow by Long Short Term Memory Recurrent Neural Network", 2015 IEEE
International Conference on Smart City/SocialCom/SustainCom (SmartCity), pp. 153-158, 2015.
[9] R. Fu, Z. Zhang and L. Li, "Using LSTM and GRU neural network methods for traffic flow prediction", Youth Academic Annual
Conference of Chinese Association of Automation (YAC), pp. 324-328, 2016.
[10] TUİK, Adrese Dayalı Nüfus Kayıt Sistemi Sonuçları 2013, Türkiye İstatistik Kurumu Haber Bülteni, Sayı: 37210, 04.Şubat.2021.
[11] TUİK, “İllere göre motorlu kara taşıtları sayısı”, Available: [Link]
December-2020-37410, [Accessed Aralık, 2020].
[12] A Liaw, M. Wiener, 2002, Classification And Regression ByRandom Forest, R News, Vol.2/3, December.
[13] L. Breiman, “Random forests”, Machine Learning, Volume 45, pp. 5-32, 2001.
[14] N. Kriegeskorte, T. Golan, Neural network models and deep learning, Current Biology, 29(7), R231–R236, 2019.
[15] L. Deng and D. Yu, “Deep Learning: Methods and Applications,” Found. Trends® Signal Process., vol. 7, no. 3–4, pp. 197–387,
2014.
[16] K. Chakraborty, K. Mehrotra, C. K. Mohan and S. Ranka, “Forecasting The Behavior of Multivariate Time Series Using Neural
Networks”, Neural Networks 5(6):961-970, 1992.
[17] Y. Duan, Y. L.V. and F. Wang, "Travel time prediction with LSTM neural network," 2016 IEEE 19th International Conference on
Intelligent Transportation Systems (ITSC), 2016, pp. 1053-1058, doi: 10.1109/ITSC.2016.7795686, 2016.
[18] S. Samui, I. Chakrabarti, S.K. Ghosh, “Tensor-Train Long Short-Term Memory for Monaural Speech Enhancement” arXiv preprint
arXiv:1812.10095, 2018
[19] F. A. Gers, J. Schmidhuber, F. A. Cummins, “Learning to forget: Continual prediction with LSTM” Neural Computation, 12 (10),
pp. 2451-2471, 2000.
[20] Ulaşım Daire Başkanlığı, “Saatlik Trafik Yoğunluk Veri Seti, Ağustos 2020 Trafik Yoğunluk Verisi” 13 Aralık, 2020.

143

View publication stats