New Speed Limit Recognition System: Software and Hardware

Validation

Nesrine Triki

1,2 a

, Mohamed Karray

and Mohamed Ksantini

National School of Engineers of Sfax, CEM Lab, University of Sfax, Sfax, Tunisia

ESME, ESME Research Lab, Ivry Sur Seine, France

Keywords: Advanced Driver Assistance Systems (ADAS), Automated Driving Systems (ADS), Speed Limit Recognition

System (SLRs), Artificial Intelligence, Belief Functions, Ensemble Learning, Embedded Systems.

Abstract: Recent advancements in intelligent driving have led to the integration of various automated systems into

vehicles, including Speed Limit Recognition systems, which play a crucial role in enhancing road safety and

autonomous driving technologies. This paper presents a comprehensive approach to Speed Limit Recognition,

based on three modules: detection, classification, and the fusion of machine learning and deep learning

classifiers. The proposed approach achieves impressive results, with an accuracy of 99.98% using Dempster

Shafer theory and 99.96% with the voting technique. The system's performance is rigorously evaluated

through simulation and hardware validation using a Raspberry Pi 4 board. Experimental results indicate high

performance rates across nine classes from the German Traffic sign Recognition Benchmark dataset in an

average processing time of 0.15 seconds.

1 INTRODUCTION

Research and development in intelligent driving have

been the subject of numerous projects and efforts in

recent years, thanks to the considerable improvement

in the performance of on-board vehicle equipment.

These advancements have allowed automotive

manufacturers to integrate systems offering various

levels of autonomy and safety into their new vehicles,

such as Advanced Driver Assistance Systems

(ADAS) and Automated Driving Systems (ADS),

both of which fall under the umbrella of automated

driving (Wintersberger et al., 2016). Among these

devices are fatigue detection systems, accident and

pedestrian detection systems, and systems for Traffic

Sign Recognition (TSR).

Speed Limit Recognition (SLR) system has made

significant advancements in recent years to enhance

road safety. In fact, it uses cameras, image

processing, and AI techniques to detect and classify

Speed Limit (SL) signs, either to assist drivers or to

take control of the vehicle. In order to ensure the

detection and classification of SL signs, a multitude

https://orcid.org/0000-0002-2770-2526

https://orcid.org/0000-0001-7293-8696

https://orcid.org/0000-0002-9928-8643

of methods based on color, shape, color and shape,

Machine Learning (ML) and Deep Learning (DL)

algorithms are used. Nevertheless, this system faces

diverse limitations, including weather conditions,

poor lighting, sign occlusion, variability in

nomenclature, etc. (Miyata, 2017). In order to

overcome these challenges, computer vision

algorithms must be robust and able to accurately

identify SL signs across a wide range of scenarios.

For these reasons, this paper proposes a new real-

time SLR system based on ML and DL techniques.

Relevant related works are presented in the second

section, and a detailed explanation of the recognition

process is provided in the third section. In the fourth

section, software and hardware validation are done in

order to confirm the obtained performance. Finally, a

conclusion and some perspectives are proposed.

2 RELATED WORKS

SLR system comprises two main components: the

detection of signs within the image, followed by the

368

Triki, N., Karray, M. and Ksantini, M.

New Speed Limit Recognition System: Software and Hardware Validation.

DOI: 10.5220/0012351000003636

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

In Proceedings of the 16th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2024) - Volume 3, pages 368-375

ISBN: 978-989-758-680-4; ISSN: 2184-433X

recognition of the speed limit value indicated on the

SL sign during the classification step. In this section,

several studies focusing on a multitude of methods

based on color, shape, ML, and DL techniques are

presented in order to recognize SL signs.

In fact, (Agudo et al., 2016) propose a real-time

framework for detecting and recognizing SL signs in

railway networks. Two different Support Vector

Machines (SVMs) are trained, the first recognizes

sign types and the second recognizes numbers. The

framework achieved a recall rate of approximately

95% on railway videos. A modified version of the

Histogram of Oriented Gradients (HOG) is used with

SVM by (Mammeri et al., 2013) to detect and

recognize over 94% of North American SL signs.

(Kundu and Mackens, 2015) use shape and intensity

information to detect American SL signs after

identifying ROIs as extremely stable extreme regions

(MSERs). They use the Kalman filter to track the

detected signs, considering only linear car

movements in the tracking phase. At the classification

stage, an ANN is used for recognition with an

accuracy of 98% on 12300 images. HOG and MSER

functions are also employed by (Soetedjo and

Somawirata, 2018) for SL sign detection and

classification, achieving a classification rate of

93.67% with a processing time of 10.75 ms. (Liu et

al., 2012) combined log-polar mapping and Locality-

constrained Linear Coding (LLC) to recognize speed

limit signs, achieving an accuracy of 97.31% on the

German Traffic Sign Recognition Benchmark

(GTSRB) dataset. An illumination-robust method for

real-time SL sign using Modified Census Transform

(MCT) and SVM, resulting in a detection rate of

99.8% and a classification rate of 98.4%.

Neural networks are used by (Miyata, 2017) for

the classification of SL signs. Convolutional Neural

Network (CNN) is used by the study (Yan et al.,

2017) in the classification stage, and the authors

achieved a recognition rate of approximately 98.51%.

In addition, (Li et al., 2016) also use CNN for the

detection and classification of SL signs by applying

pre-processing and post-processing to images in order

to improve performance, achieving an accuracy of

over 97% on the LISA-TS dataset. These studies

showcase various approaches and techniques,

including SVMs, HOG, MSER, CNNs, etc.

Unfortunately, these methods exhibit certain

limitations including challenges in generalizing to

diverse conditions, dependency on training data

quality, computational resource requirements,

potential for false positives and negatives, sensitivity

to environmental factors, adaptability to different sign

designs, real-time processing constraints, etc. To

address these limitations and improve SLR

performance, a novel methodology based on fusion

techniques combining ML and DL classifiers is

proposed. This approach will be further explored in

the subsequent section.

3 PROPOSED SPEED LIMIT

RECOGNITION APPROACH

SLR involves three main modules: the first detects

speed limit signs (Speed Limit Detection, SLD), the

second classifies the detected signs (Speed Limit

Classification, SLC), and the last merges pre-trained

classifiers (Speed Limit Classifiers Fusion, SLCF).

The recognition process begins with capturing images

using a camera and applying preprocessing steps. The

pre-processed image is sent to the SLD module for

detection using the Haar Cascade method, a ML

object detection technique introduced by (Viola and

Jones, 2001). It locates swiftly potential Region of

Interests (ROI) within captured images. The SLC

module employs a new developed CNN model (Deep

Speed Limit, DeepSL) trained on SL images from

GTSRB dataset. In order to improve classification, k-

Nearest Neighbors (KNN), Random Forest (RF), and

SVM are used and combined in the SLCF module

using Ensemble Learning (EL) or Dempster Shafer

(DS) theory aiming to enhance the recognition

process by finding the best combination.

3.1 Speed Limit Detection Module

In the context of detecting SL road signs images, the

Haar Cascade method is particularly used for its

effectiveness in real time object detection. In fact, this

method operates by using a set of simple rectangular

features called Haar-like features computed at various

scales and positions across the input image, serving

as templates that capture different characteristics of

the object under consideration, such as edges,

corners, or distinctive patterns. The integral image

technique is applied to efficiently calculate these

features, contributing to a streamlined computational

process. Subsequently, AdaBoost, a cascade ML

classifier is trained to identify a small yet crucial

subset of features capable of effectively

distinguishing between positive and negative

samples. This sequential approach enables the

method to swiftly discard image areas unlikely to

contain the target object. This robust detection system

is then applied to pre-processed images to accurately

locate and extract speed limit signs.

New Speed Limit Recognition System: Software and Hardware Validation

369

3.2 Speed Limit Classification Module

The use of DL models such as Convolutional Neural

Networks (CNNs), Deep Neural Networks (DNNs),

and others has garnered significant interest in both the

academic and industrial communities due to their

high performance in SL classification compared to

traditional ML classifiers using various image

datasets (Triki et al., 2023). In this paper, SL road

signs classes from the GTSRB are used.

3.2.1 Speed Limit Road Signs Dataset

Nine SL Traffic signs classes (20, 30, 50, 70, 80, end

of 80, 100, and 120 km/h) from the GTSRB dataset

are used, including approximately 13200 images

captured under various environmental and weather

conditions. These classes exhibit imbalanced

distributions. In fact, several approaches can be

employed to address this issue. A study conducted by

(Rout et al., 2018) explored the use of data

augmentation, which generates additional samples for

classes by applying various transformations such as

rotation, translation, scaling, or adding noise. This

technique diversifies the training set, mitigates

overfitting, and ultimately improves generalization

performance. Furthermore, oversampling techniques

increase the number of samples in the minority class

through replication or synthetic generation, while

undersampling techniques decrease the number of

samples in the majority class. In order to address this

issue, data augmentation, over-sampling, and under-

sampling techniques are applied to the training image

set.

Before addressing the class imbalance problem in

the data used for training DeepSL, a preprocessing

step is required. Initially, images are converted to

grayscale. Then, histogram equalization is applied to

enhance the overall image contrast. Subsequently,

image normalization and reshaping are performed.

3.2.2 DeepSL Classification Model

DeepSL, a new ConvNet has the architecture detailed

as follows: detected SL grayscale signs, are passed

through an initial Conv2D layer with 32 filters of size

(3x3), followed by a ReLU activation function. This

layer extracts features, such as edges and textures. A

Batch Normalization layer follows the first

convolutional layer, helping to normalize activations

and stabilize network training. Next, a second

Conv2D layer with the same parameters as the first

convolutional layer is added, followed by another

Batch Normalization layer. This sequence of

convolutional and normalization layers is repeated a

second time with 64-sized filters. Between each pair

of convolutional layers, a 2x2 Max pooling 2D layer

is added to extract the most important features from

the previous layer. Dropout layers with a rate of 0.25

are added after each Max pooling layer to prevent

overfitting. Once all features have been extracted, a

Flatten layer is used. Subsequently, a Dense layer

with 512 neurons and a ReLU activation function is

added to perform a linear combination of the

previously extracted features using connection

weights, A Batch Normalization layer, and a Dropout

layer with a rate of 0.5 are added. Finally, a last Dense

layer with nine neurons and a SoftMax activation

function is added to perform the input images

classification into the specified SL classes from

GTSRB dataset.

The DeepSL model is trained for 70 epochs with

a batch size of 32 on a GPU. Results of performance

of the model after applying these three techniques are

represented in Figure 1, Figure 2, and Figure 3.

Figure 1: Accuracy and loss curves of the training and

testing sets using data augmentation technique.

Figure 2: Loss curves of the training and testing sets using

over-sampling technique.

Figure 3: Loss curves of the training and testing sets using

under-sampling technique.

Examining Figures 1 and 2, the loss curves

demonstrate a remarkable fit of the DeepSL model

when using data augmentation and oversampling,

unlike those in Figure 3, which employ

ICAART 2024 - 16th International Conference on Agents and Artiﬁcial Intelligence

370

undersampling and exhibit less significant

performance. The precision (P), recall (R), and F1

weighted average scores are calculated in order to

further evaluate achieved performances. Obtained

weighted average scores are summarized in Table 1.

Table 1: Obtained weighted average scores.

Handling

Techniques P (%) R (%) F1 (%)

Data Au

mentation 99.81 99.81 99.81

Ove

-sam

lin

99.69 99.69 99.69

Unde

-sam

lin

96.37 96.27 96.27

The F1 score, a harmonic mean of precision and

recall, is used to assess DeepSL performance of the

model. Based on the results presented in Figures 1

and 2 and the weighted F1 scores from Table 1, the

performance of the DeepSL using data augmentation

and oversampling techniques is very close. Therefore,

an evaluation of the model on a real test set is

necessary to select the best model. An example of

images from the test set is illustrated in Figure 4.

Figure 4: Examples of SL images from the Test set.

By applying the DeepSL model with data

augmentation, the model achieves an F1 score of

99.19%, whereas it is approximately 95.22% when

using oversampling on the data. Consequently, the

DeepSL model using data augmentation is chosen for

its superior classification performance. In addition to

the F1 score, other metrics are used, such as the

confusion matrix and the classification report

illustrated in Figures 5 and 6.

Figure 5: Confusion Matrix of the DeepSL model with data

augmentation on the test set.

Figure 6: Classification report of the DeepSL model with

data augmentation on the test set.

Indeed, in this set of images, only one error was

observed outside the main diagonal of the test

confusion matrix, where a speed limit sign of 80 km/h

was predicted as a 30 km/h speed limit sign.

Moreover, excellent results are obtained for the

classification of SL signs by class. The weighted F1

scores are 98% for 2 classes and 100% for the

remaining 7 classes. In general, prediction errors can

be attributed to various factors, including the quality

and quantity of images used for training or the

classifier itself.

3.3 Speed Limit Traffic Sign Classifiers

Fusion Module

Various methodologies have been applied for image

recognition and have indeed produced good results.

However, they suffer from the loss of details during

feature extraction due to noise in the image, the

presence of similar objects or complex backgrounds,

variations in lighting, etc. To address these issues,

fusion methods combining the results of multiple

classification algorithms are employed such as DS

theory and EL.

In fact, DS theory is a formalism for making

decisions in uncertain situations (Dempster, 2008). It

is based on concepts including:

 The mass function (m) is used to represent

uncertain or incomplete information about

hypotheses in a problem. It is defined by:

[]

()

m: 2 0 ,1 with m A 1

A⊆Ω

→=



Ω

(1)

 The new mass function of the attenuation

operation defined by equation (2) is applied.

() ()

m A µ * m A ; A =∀≠Ω

(2)

 Information fusion is done through the DS

fusion rule by calculating a global mass

function defined by:

() () ()

m(m1 m2 A)C m1 * 2B

AB C A B=∩

⊕=



(3)

New Speed Limit Recognition System: Software and Hardware Validation

371

 The pignistic transformation converts a mass

function into a pignistic probability

measure. It is calculated using:

()

{A , ω A}

)

((

)

pω

1m A

⊆∈

−∅



Ω

(4)

 The decision will be made by choosing the

element x with the highest probability from

the pignistic transformation using:

() ()()

Rp x argmax Betp ω x

∈

Ω

(5)

In addition to DS theory, EL, a powerful ML

technique combining multiple classifiers using

various methods to produce more accurate and

reliable final decision is used (Mohammed et al.,

2023). The most common EL techniques include

bagging, boosting, stacking, and voting. The Bagging

involves creating multiple copies of the same model

and training each copy in parallel on random subsets

of the dataset. The Boosting sequentially trains

multiple relatively weak models. Each model is

responsible for correcting the errors of its

predecessor. The Stacking aims to predict the best

classifiers and assign weights to them. The Voting

takes advantage of the performance of numerous

models, making it less susceptible to significant

errors or misclassifications from a single model.

There are two types of voting: Hard Voting (HV) and

Soft Voting (SV). In the following, the DS theory and

the soft voting technique are applied.

3.3.1 Application of Classifiers Fusion Using

DS Theory

First, SVM, KNN, and RF classifiers are trained on

SL signs (13200 images: 75% for training and 25%

for testing) from the GTSRB dataset, using three

feature descriptors (FD): RGB color descriptor

(FD1), 3D color histogram (FD2), and HOG

descriptor (FD3). F1 scores are presented by Table 2

Table 2: F1 scores obtained by SVM, KNN, and RF.

Classification Methods

F1 (%)

FD ML Classifiers

FD1

SVM 34.85

KNN 79.39

RF 74.36

FD2

SVM 93.33

KNN 92.67

RF 96.42

FD3

SVM 95.88

KNN 95.85

RF 92.58

SVM, RF, KNN, and DeepSL classifiers are then

fused using the DS theory. The results of the different

classifier combinations are presented in Table 3.

Table 3: Classification results using the DS theory.

Combined classifiers F1

(

)

KNN and SVM 99.31

RF and SVM 99.35

RF and KNN 94.48

RF and DeepSL 99.38

KNN and DeepSL 99.29

SVM and Dee

SL 93.3

KNN and RF and Dee

SL 99.93

SVM and KNN and RF 99.91

SVM and KNN and DeepSL 99.91

SVM and RF and DeepSL 99.92

4 SVM and RF and KNN and Dee

SL 99.98

3.3.2 Application of Fusion Using Voting

Technique from Ensemble Learning

Initially, features from training images are extracted

using the DeepSL model in order to capture important

features which are then used separately to train KNN,

RF and SVM classifiers. Table 4 summarizes F1

scores of KNN, RF, and SVM on the testing set

described by Figure 4.

Table 4: Weighted F1 Score of the KNN, RF, and SVM.

SL classification methods Wei

hted F1 Score

(

)

KNN and Dee

SL 99.88

RF and DeepSL 99.90

SVM and DeepSL 99.87

According to Table 5, the ML fusion classifiers

using the SV improves significantly F1 score

compared to using each classifier separately. In fact,

the SV collects predicted probabilities for each class

label and predicts it with the highest probability.

Table 5: F1 scores of the fusion of ML classifiers using

DeepSL.

Combined Classifiers

F1 Soft Voting (%)

KNN and SVM

99.90

RF and SVM

99.87

RF and KNN

99.96

3 SVM and KNN and RF

99.90

3.4 Comparative Study of SLC

Approaches

Several studies on SLC have been presented.

(Soetedjo and Somawirata, 2018) present a method

for speed limit sign classification using features such

ICAART 2024 - 16th International Conference on Agents and Artiﬁcial Intelligence

372

as HOG and Maximally Stable Extremal Regions

(MSER) and achieved a classification rate of

93.67%.Another study introduces a speed limit sign

classification technique based on the HOG and ring

areas by (Soetedjo et al., 2017). The method divides

an image into multiple rings and computes the HOG

feature on each ring. In the matching process, a

weight is assigned to each ring to calculate the HOG

feature distance between the test image and the

reference image. Experimental results show that the

proposed algorithm achieves a high classification rate

of 97.8%. Regarding the study realised by (Saadna et

al., 2019), a two-SVM cascade architecture are

designed in the classification phase. The first SVM is

trained on the GTSRB dataset to determine whether

the detected region is a speed limit sign or not, and

the second SVM is trained on the MNIST dataset to

recognize the value of speed limit signs. The system

achieves a precision of 99.08% on the GTSRB

dataset. Comparing various obtained results through

DS theory and SV fusion, KNN and RF fusion using

DeepSL as a feature extractor are the best, achieving

a classification rate of 99.98% and 99.96%

respectively. Furthermore, this result surpasses the

performance of the other studies mentioned earlier.

To confirm these performances in real-world

scenarios, a validation of the proposed approach is

realised in the next part.

4 IMPLEMENTATION AND

VALIDATION OF THE

PROPOSED SLR SYSTEM

4.1 Software Validation

Software validation is an essential step to confirm the

efficiency and reliability of the proposed SLR system.

In this section, the system is validated using a car

simulator in addition to computer and road scenes.

4.1.1 Simulator Validation

A simulator is a tool that provides interactive virtual

environments similar to real life to simulate situations

in various fields, including the automotive industry.

The choice has been to enhance an open-source

simulator called Udacity (Du et al., 2019), initially

containing only the map, by incorporating various

signs and signals such as 30 km/h, 50 km/h, 60 km/h,

70 km/h, and 80 km/h to test the proposed solution.

In order to validate the SLR system through the

car simulator, two steps must be completed: the

training mode generating a trained CNN model for

autonomous driving and the autonomous mode

testing the effectiveness of the SLR system in real

scenarios. An example of simulating this driving

mode is depicted in Figure 7.

Figure 7: Example of recognized sign using the simulator

in autonomous mode.

A test is conducted on the simulator, and its

performance results are collected in Table 6. SLR

system correctly recognizes (CR) various signs which

enhances reliability and provides reassurance for any

potential real tests.

Table 6: Results of SLR system via the simulator.

SL road signs

CR / Total signs 3/3 3/3 1/1 1/1 1/1

4.1.2 Validation by Simulation

Validation by simulation through driving sequences

on urban roads or highways is a common approach

for testing and validating recognition systems,

especially those related to road signs. Indeed, this

type of validation allows for simulating different

environmental driving conditions and evaluating the

recognition system's performance in various

scenarios. In order to validate the proposed SLR

system through simulation, a video sequence

describing a road scene, rich in speed limit signs,

lasting approximately three minutes, is used.

Figure 8: Examples of SL signs correctly recognized by the

SLR system.

Simulation is first performed using a PC (Intel®

Core (TM) i5-7200 CPU, 64-bit, 8 GB RAM) and

Google Colab with 12.4 GB of RAM. To evaluate the

system's performance, recognition time and

classification rate are calculated. The system achieves

an average of 0.06s to identify each detected sign in

the case of PC simulation and an average of 0.025s

New Speed Limit Recognition System: Software and Hardware Validation

373

using Colab. Figure 8 shows examples of correctly

recognized SL signs by SLR system.

4.2 Hardware Validation

The hardware architecture of a system varies

depending on its specific processing and performance

requirements. Indeed, there are different hardware

architectures based on a CPU (Central Processing

Unit), GPU (Graphics Processing Unit), FPGA

(Field-Programmable Gate Array), or heterogeneous

architectures that combine different types of units to

leverage their specific advantages (Hu et al., 2022).

In order to validate an architecture on a hardware

target, the performance of the core used for image

processing (execution time and accuracy), available

memory and its type for efficient resource use, and

the availability of libraries and development tools to

facilitate implementation, testing, and future

improvements must be taken into account. Based on

a study of the characteristics of the different types of

boards, the validation and evaluation of the SLR

system are carried out on the Raspberry Pi 4. Indeed,

this choice is based on its technical specifications and

its adaptability for artificial intelligence applications.

In fact, featuring a quad-core ARM Cortex-A72

processor, the used Raspberry Pi 4 board provides

enhanced processing capabilities with 4GB of RAM.

Moreover, it is configured with the necessary

software and with the appropriate image processing

and ML and DL libraries such as OpenCV,

TensorFlow, Keras, etc. necessary for the proper

functioning of the system. Once the system is

installed, the same driving sequence used for software

validation is reused to perform hardware validation in

order to be faithful to the real-world environment. In

this step, the camera is positioned facing the screen to

view the video sequence. Tests are conducted to

assess the SLR performance, including processing

speed and the recognition rate of road signs. Obtained

results are summarized in Table 7.

Table 7: Evaluation results of the SLR system on Raspberry

Pi 4 board.

Results

Raspberry Pi

Correctl

reco

nized SL si

ns 18/20

Unrecognized SL signs 0/20

Incorrectly recognized SL signs 2/20

Average SL sign recognition spee

0.15s (6 ips)

Some work related to the SLR system has

undergone experiments on the Raspberry Pi board.

(Akshay et al., 2018) develop a SLR system using a

Raspberry Pi, focusing on SL signs, and considering

the stability of color detection under varying daylight

conditions. The results show that their system

achieves an accuracy of 80% with processing times

of up to 2s. Furthermore (Isa et al., 2022) implement

a real-time SLR system using the Raspberry Pi 3

board with ML algorithms to identify sign types and

send alerts to the driver, considering 5 different sign

classes. Results show that the average accuracy of

sign recognition across the five classes is above 90%,

and the maximum average time to determine the sign

type in the system is 3.44 s when the car is traveling

at 50 km/h. The proposed SLR system surpasses the

mentioned approaches in terms of precision

concerning the number of classes, with a score of

90% (18 well

recognized signs /20) for 9 classes and in

terms of processing time, with an average of 0.15 s.

5 CONCLUSIONS

The design and implementation of autonomous

vehicles are fields of research that are constantly

evolving. SLR system is an important component for

ADAS and ADS, given its considerable contribution

to user comfort, improved road safety, and adherence

to traffic rules. To achieve reliable recognition,

several parameters and constraints must be taken into

account, including environmental conditions and

response time. Indeed, signs may exhibit variations in

their appearance, degradation, and partial obstruction.

Moreover, this system must be fast and efficient in

sign detection, classification, and interpretation to

make appropriate decisions within timely intervals.

This paper proposes a comprehensive SLR

approach, covering a detection module based on Haar

Cascade technique, a classification module

employing a new model (DeepSL), and a fusion

module using the DS theory and EL. Obtained results

are very satisfactory. In fact, the classification rates

reach 99.98% and 99.96%, respectively, for DS

theory and the voting technique. The proposed

approach is rigorously evaluated through simulation

and hardware validation on the Raspberry Pi 4 board,

demonstrating promising results in terms of accuracy

and processing time, achieving a correctly

recognizing 18 out of 20 road signs images, across 9

different SL classes (from the GTSRB dataset) with a

processing time of 0.15 s. Ultimately, this research

significantly contributes to the improvement of the

driver road safety and the transportation efficiency by

providing valuable insights for the implementation of

an SLR system. As a continuation of this work, we

propose to expand the SLR system to recognize a

wider range of sign categories from different

ICAART 2024 - 16th International Conference on Agents and Artiﬁcial Intelligence

374

countries in order to improve recognition in various

contexts. Additionally, we contemplate hardware

validation using various hardware architectures like

SoC and Nano Jetson boards for real-world testing.

REFERENCES

Agudo, D., Sánchez, Á., Vélez, J.F., Belén Moreno, A.,

2016. Real-time railway speed limit sign recognition

from video sequences, in 2016 International

Conference on Systems, Signals and Image Processing

(IWSSIP).

Akshay, G., Dinesh, K., Scholars, U., 2018. Road sign

recognition system using raspberry pi.” International

Journal of Pure and Applied Mathematics 119, 15 ,

1845–1850.

Bi, Y., 2012. The impact of diversity on the accuracy of

evidential classifier ensembles. International Journal of

Approximate Reasoning 53 4 , 584–607.

Brown, J.B., 2018. Classifiers and their Metrics Quantified.

Molecular Informatics 37 1–2 , 1700127.

Dempster, A., 2008. Upper and Lower Probabilities

Induced by a Multivalued Mapping. pp. 57–72.

Du, S., Guo, H., Simpson, A., 2019. Self-Driving Car

Steering Angle Prediction Based on Image Recognition

Hu, N., Wang, C., Zhou, X., 2022. FLIA: Architecture of

Collaborated Mobile GPU and FPGA Heterogeneous

Computing. Electronics 11 22 , 3756.

Isa, I.S.B.M., Yeong, C.J., Azyze, N.L.A. bin M.S., 2022.

Real-time traffic sign detection and recognition using

Raspberry Pi. International Journal of Electrical and

Computer Engineering (IJECE) 12 1 , 331–338.

Kundu, S., Mackens, P., 2015. Speed Limit Sign

Recognition Using MSER and Artificial Neural

Networks.

Li, Y., Mogelmose, A., Trivedi, M.M., 2016. Pushing the

“Speed Limit”: High-Accuracy US Traffic Sign

Recognition With Convolutional Neural Networks.

IEEE Transactions on Intelligent Vehicles.

Liu, B., Liu, H., Luo, X., Sun, F., 2012. Speed Limit Sign

Recognition Using Log-Polar Mapping and Visual

Codebook.

Mammeri, A., Boukerche, A., Feng, J., Wang, R., 2013.

North-American speed limit sign detection and

recognition for smart cars, in: 38th Annual IEEE

Conference on Local Computer Networks - Workshops.

pp. 154–161.

Minary, P., Pichon, F., Mercier, D., Lefevre, E., Droit, B.,

2017. Face pixel detection using evidential calibration

and fusion. International Journal of Approximate

Reasoning 91, 202–215.

Miyata, S., 2017. Automatic Recognition of Speed Limits

on Speed-Limit Signs by Using Machine Learning.

Journal of Imaging 3 3 , 25.

Mohammed, A., Kora, R., 2023. A comprehensive review

on ensemble deep learning: Opportunities and

challenges. Journal of King Saud University -

Computer and Information Sciences 35 2 , 757–774.

Rout, N., Mishra, D., Mallick, M., 2018. Handling

Imbalanced Data: A Survey. pp. 431–443.

Saadna, Y., Behloul, A., Mezzoudj, S., 2019. Speed limit

sign detection and recognition system using SVM and

MNIST datasets. Neural Computing and Applications

31.

Soetedjo, A., Somawirata, I., 2017. Circular traffic sign

classification using hogbased ring partitioned

matching. International Journal on Smart Sensing and

Intelligent Systems 10, 735–753.

Soetedjo, A., Somawirata, I.K., 2018. Speed Limit Traffic

Sign Classification Using Multiple Features Matching,

in: Electrical Engineering. Springer, Singapore, pp.

210–217.

Triki, N., Karray, M., Ksantini, M., 2023. A Real-Time

Traffic Sign Recognition Method Using a New

Attention-Based Deep Convolutional Neural Network

for Smart Vehicles. Applied Sciences 13 8 , 4793.

Triki, N., Ksantini, M., Karray, M., 2021. Traffic Sign

Recognition System based on Belief Functions Theory,

in: Proceedings of the 13th International Conference on

Agents and Artificial Intelligence, SCITEPRESS, pp.

775–780.

Viola, P., Jones, M., 2001. Rapid object detection using a

boosted cascade of simple features, in: Proceedings of

the 2001 IEEE Computer Society Conference on

Computer Vision and Pattern Recognition, USA, p. I-

511-I–518.

Wintersberger, P., Riener, A., 2016. Trust in Technology as

a Safety Aspect in Highly Automated Driving. i-com 15

3 , 297–310. doi:10.1515/icom-2016-0034

Xu, P., Davoine, F., Zha, H., Denœux, T., 2016. Evidential

calibration of binary SVM classifiers. International

Journal of Approximate Reasoning, BELIEF 2014 –

Third International Conference on Belief Functions 72,

55–70.

Yan, G., Ming, Y., Shi, S., Feng, C., 2017. The recognition

of traffic speed limit sign in hazy weather. Journal of

Intelligent & Fuzzy Systems 33, 1–11.

New Speed Limit Recognition System: Software and Hardware Validation

375