Deep Learning-Based Algorithms in Solving Traffic Jam in Smart

Transportation

Jingru Deng

Data Science and Big Data Technology, South China University of Technology, Guangzhou, China

Keywords: Traffic Flow Prediction, Machine Learning, Smart Transportation, Artificial Neural Network.

Abstract: With the increase in the number of private cars, people's travel has become more convenient. However,

currently artificial intelligence has not played a particularly significant role in solving traffic congestion.

Traffic congestion has lots of disadvantage: it can increase people's commuting time; it can increase travel

costs and so on. This article aims at providing an overview of some possible ways in predicting traffic flow,

ranging from machine-learning based method to deep learning methods, which give a feasible scheme for the

traffic system to moderate the traffic light time and better smooth the traffic flow. In the discussion part, the

article analysis that machine-learning based methods still have shortcomings in terms of interpretability and

adaptability. In the future, the predicting method will be improved by adapting SHapley Additive exPlanations

and domain adaptation. Also, computational speed also needs to be taken into account, with the use of parallel

computing. The author proposes a framework that focuses on parallel computing. This article provides a good

overview of the field of predicting traffic flow.

1 INTRODUCTION

With the rapid development of Internet of Things

(IOT), transportation facilities in the city are

connected to the traffic network, giving rise to the

concept of smart transportation. Smart transportation

is considered a general term that encompasses route

optimization, parking, street lighting, accident

prevention and detection, road anomalies, and

infrastructure applications (Zantalis, 2019). Some

apps e.g. Gaode in the field of transportation, it can

help users optimize the route to destination. They can

easily obtain the location information of a certain

vehicle and predict the time it will take to reach a

certain station by installing mobile devices on public

transportation. Meanwhile, the increase in the number

of cars puts pressure on traffic system. In the urban

area, it is often easy to encounter traffic congestion at

intersections due to excessive traffic flow or

accidents. Traffic jams are detrimental to people’s

lives in many aspects. It takes more time for people

to wait and consumes more gasoline. Thus, how to

solve traffic congestion in a better way is what needs

to be considered.

https://orcid.org/0009-0009-2052-1122

Artificial intelligence technology includes data

excavation and intelligent algorithms, which means

that it can search for useful information through its

internal algorithm and can solve certain types of

issues. In real situation, collecting information on

vehicles can be regarded as data excavation, while

getting a solution for traffic jams is what the artificial

algorithm can do. Therefore, artificial intelligence

can be well integrated with intelligent transportation,

especially, the prediction of traffic flow.

In the aspect of predicting the traffic flow through

artificial intelligence, many researchers have carried

out some relative works. In early research,

researchers employed mathematical statistics or

traditional machine learning method to build traffic

flow prediction model. For example, Wang upgraded

the Cellular Automata model suitable for describing

the spread of traffic congestion by taking certain set

called “Squeezing to change lanes” into consideration

(Wang, 2010). Castro et al. utilized the advantages of

online learning methods to predict traffic flow under

different conditions, and thus proposed a supervised

statistical learning technique called Online

Regression Support Vector Machine (Castro-Neto,

2009). However, the conventional machine learning

Deng, J.

Deep Learning-Based Algorithms in Solving Trafﬁc Jam in Smart Transportation.

DOI: 10.5220/0012958500004508

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 1st International Conference on Engineering Management, Information Technology and Intelligence (EMITI 2024), pages 541-545

ISBN: 978-989-758-713-9

541

method has various disadvantages. Traditional

machine learning method requires people to establish

eigenvalues and has a significant workload in

identifying feature values. Moreover, it is impossible

to exhaustively list all the characteristics. With the

appearance of neural network, deep learning can find

out the rules without establishing eigenvalues for

them. Thus, researchers tried using deep learning to

tackle the prediction issue. Liang et al. improved the

accuracy of predicting urban traffic flow through

deep learning, specifically the Spatio-Temporal

Relation Network (STRN), which consists of

Convolutional Neural Network (CNN) (Liang, 2021).

Mao et al. established a platform based on Spark and

Hadoop to achieve real-time monitoring of urban

traffic flow. This platform also has some other

functions, such as the ability to analyze specific types

(Mao, 2022). Due to the rapid development of this

field and its importance, it is necessary to make a

comprehensive overview of this direction.

The rest of the paper is constructed as follows. In

Section 2, The article will provide a thorough

explanation of the procedures used by others in their

workflow, such as traditional machine learning

method and Artificial Neural Network (ANN). In

section 3, the author will discuss shortcomings of

current researches and future development directions.

Finally, section 4 sums up the paper.

2 METHOD

2.1 Framework of Machine

Learning-Based Methods in

Predicting Traffic Flow

Machine learning process for traffic flow prediction

shown in Figure 1 can be loosely broken down into

six steps. The first step is data collection, which

obtains various kind of information e.g. weather,

historical traffic flow, time, holidays, events. Then, in

the data preprocessing progress, data will experience

steps like cleaning, normalization for the reason that

different data are of different importance for

predicting traffic flow forecasts, and there might be

some abnormal values. This step aims at enhancing

the quality of data. In the next step, a model that can

predict the traffic flow is constructed, consisting of

suitable algorithms. Algorithms such as Random

Forest, K-NN are prevailing in traditional machine

learning. In the fourth step, when the model is trained,

whose parameters is dynamically moderating in order

to get a more precise prediction value of the traffic

flow. In the fifth step, when the model is well trained,

it will be accessed by some index like some metrics

to figure out whether the model’s parameter or

algorithm need changing or not. Finally, the model

can be deployed and help police visualize where

traffic jam will take place. The section left describes

the details of model building and training.

Figure 1: The workflow of machine learning-based

methods in predicting traffic flow (Photo/ Picture

credit: Original).

2.2 Machine Learning Algorithms

2.2.1 Random Forest-Based Prediction

When it comes to model building and training

progress, Chen et al. proposed Random Forest

Prediction model shown in Figure 2 for traffic flow

prediction (Chen, 2018). Random Forest is a

combined algorithm that uses multiple tree learners

for classification and regression prediction. The

sample to be classified is denoted as Xt=Xt1, Xt2,

Xt3, where Xt1, Xt2, and Xt3 are mapped to the

predicted traffic flow, velocity, and occupancy of

target section in the upcoming period. Icons of

different shapes and colors represent distinct traffic

situations (Chen, 2018). When a sample to be

classified is given, each decision tree calculates the

percentage of different classes in the training sample

at the leaf node where the sample falls to determine

the classification result of the sample (Zhou, 2019).

Finally, the classification results from all the decision

trees are collected and voted. The one which has the

highest number of votes is the prediction of the traffic

state.

Figure 2: The framework of Random Forest Prediction

Model for traffic flow prediction (Photo/Picture credit:

Original).

2.2.2 K-Nearest Neighbour Classification

(K-NN)

K-NN shown in Figure 3 is another algorithm that can

be applied to build the model. K-NN requires 5

EMITI 2024 - International Conference on Engineering Management, Information Technology and Intelligence

542

elements: Complete historical datasets, state vector,

distant standard, K nearest neighbour matching and

prediction algorithm (Cover, 1967). The first step is

to construct a historical database from data that has

been preprocessed. The historical database needs to

be large enough and has representative data. The

second step is to define the state vector. Next, select

a distance metric criterion. Then, search for K

neighbours of the current vector and put it into

prediction algorithm to get a prediction result. The

prediction algorithm can be divided into 2 parts:

arithmetic mean prediction algorithm and weighted

prediction algorithm (Lin, 2015).

Figure 3: The framework of K-NN for traffic flow

prediction (Photo/Picture credit: Original).

2.3 Deep Learning Methods

What the framework of deep learning methods look

like is quite similar to that of the machine learning

methods. However, since deep learning can extract

abstract features and learn automatically in the model

building progress, it is better than classical machine

learning.

2.3.1 Convolutional Neural Network

CNN is characterized by its deep feedforward neural

network that utilizes convolutional computation and

deep structure (Qiu, 2022). Ma et al. migrate the

application of CNN to the field of traffic forecasting

shown in Figure 4. In the data preprocessing process,

the traffic flow is transformed into graph. There are 2

main steps in the model building process. Firstly,

images that have spatiotemporal characteristics and

are produced by traffic network are the main resource

of the model input (Ma, 2017). Secondly, both the

convolutional layer and pooling layers, which are two

of the most crucial components of the CNN model,

extract features from the input and integrate with each

other at the same time (Ma, 2017). Finally, through a

fully connected layer, vectors representing features

are transformed into model outputs (Ma, 2017).

Figure 4: The framework of CNN for traffic flow prediction

(Photo/Picture credit: Original).

2.3.2 Spatio-Temporal Relation Network

(STRN)

STRN shown in Figure 5 is another model that can

predict traffic flow precisely (Liang, 2021). This

model collects traffic flow at different times and

selects key time steps. Then, it preprocesses the data

by dividing the data into three categories: closeness,

period and trend. These three categories are processed

in a convolutional layer. Also, external factors like

weather, POI&RN are preprocessed through Meta

Learner. Then, all of the preprocessed data are sent to

the backbone network and the GloNet for model

learning.

Figure 5: The framework of STRN for traffic flow

prediction (Photo/Picture credit: Original).

3 DISCUSSIONS

Although significant progress has been achieved in

this field, the deep learning-based model has some

drawbacks. The interpretability of deep learning is

relatively poor. It can provide people a prediction

result, but it is challenging for people to understand

why it gets such a result. Especially when error occurs

in prediction, people have no way to find out the

reason. Also, the applicability needs to be further

improved. In previous researches, the prediction

models are used in certain roads or areas. If the

circumstance changes, for example, the prediction

model needs to be applied in a more complicated

situation, it may cause error. Furthermore, although

currently there have been many traffic flow prediction

models, their computing speed is very important since

it need to predict the real-time traffic flow.

There are some ways that can be considered to

tackle the issues above in the future. Firstly, the

SHapley Additive exPlanations (SHAP) can com-

pensate for the shortcomings of poor interpretability

because it can quantify the contribution of each factor

to the prediction results and help understand the

decision process of the model. The Domain

Adaptation can tackle the applicability issue through

feature-based and instance-based approach. Finally, it

is possible to take parallel computing into

consideration to solve the computing speed problem.

Deep Learning-Based Algorithms in Solving Trafﬁc Jam in Smart Transportation

543

Figure 6: The potential framework in the future (Photo/Picture credit: Original).

A possible framework shown in Figure 6 is proposed

in this study. The model can be splited into three

layers. The first layer is data collection layer, then

comes the computational and analysis layer and the

last layer is the application layer. In the first layer, it

is required to collect external factors and past

situation and preprocess them. Then, in the

computational and analysis layer, since external

factors, closeness, period, and trend can be

considered independent of each other, parallel

computing is feasible for them. These four factors are

placed on the four computing nodes of Spark for

parallel computation and processed through

convolutional layers. However, there are many

redundant and irrelevant information in the obtained

four processing results. Therefore, matrix joint

analysis can be applied to capture key features and

extract effective information for these four results.

Input the results into Graph Transformer for parallel

computing, while in the previous research it will be

processed and analyzed by Backbone Network and

Glonet, but this method is difficult to parallelize. Next,

by using reinforcement learning models with the goal

of minimizing congestion, the optimized future traffic

light duration is obtained. The last layer is the

application layer, it depends on whether the

transportation department adopts this optimization

plan.

4 CONCLUSIONS

In this work, the article provided a review of machine

learning in traffic flow prediction. It summarizes

some traditional machine-based methods and deep

learning methods in predicting traffic flow. The

methods of this research include Random Forest, K-

NN, CNN and STRN. Through the discussion and

analysis, currently the machine-learning method still

has some disadvantages such as interpretability,

applicability and computing speed. This article has

EMITI 2024 - International Conference on Engineering Management, Information Technology and Intelligence

544

some drawbacks because some classic models are out

of consideration. In the future, further study plans to

carry out a more comprehensive job and incorporate

more algorithms related to traffic flow prediction into

the system.

REFERENCES

Breiman, L. 2001. Random forests. Machine learning, 45,

5-32.

Castro-Neto, M., Jeong, Y. S., Jeong, M. K., & Han, L. D.

2009. Online-SVR for short-term traffic flow prediction

under typical and atypical traffic conditions. Expert

systems with applications, 36(3), 6164-6173.

Chen, Z., et al. 2018. Short-term traffic state prediction

approach based on FCM and random forest. Journal of

Electronics and Information Technology, 40(8), 1879-

1886.

Cover, T., & Hart, P. 1967. Nearest neighbor pattern

classification. IEEE transactions on information theory,

13(1), 21-27.

Liang, Y., Ouyang, K., Sun, J., Wang, Y., Zhang, J., Zheng,

Y., ... & Zimmermann, R. 2021. Fine-grained urban

flow prediction. In Proceedings of the Web Conference

2021 (pp. 1833-1845).

Lin, C. 2015. Research on Short term Traffic Flow

Prediction Algorithm Based on K-Nearest Neighbor

Non parametric Regression (Doctor dissertation,

Chengdu: University of Electronic Science and

Technology of China)

Ma, X., Dai, Z., He, Z., Ma, J., Wang, Y., & Wang, Y. 2017.

Learning traffic as images: A deep convolutional neural

network for large-scale transportation network speed

prediction. Sensors, 17(4), 818.

Mao,L., Deng,S., Zhao,D., Tang,L.&Sun,X. 2022.

Application and Reflection of New Energy Vehicle

Monitoring Platform in Driving and Charging

Scenarios Journal of Electrical Engineering

Technology, 37 (1), 48-57

Qiu, Y., Wang, J., Jin, Z., Chen, H., Zhang, M., & Guo, L.

2022. Pose-guided matching based on deep learning for

assessing quality of action on rehabilitation

training. Biomedical Signal Processing and

Control, 72, 103323.

Wang,Y., 2010. A simulation study on road traffic

congestion based on cellular automata Journal of

System Simulation, (9), 2149-2154

Zantalis, F., Koulouras, G., Karabetsos, S., & Kandris, D.

2019. A review of machine learning and IoT in smart

transportation. Future Internet, 11(4), 94.

Zhou, Z. H., & Feng, J. 2019. Deep forest. National science

review, 6(1), 74-86.

Deep Learning-Based Algorithms in Solving Trafﬁc Jam in Smart Transportation

545