Proceedings of the Second International Conference on Applied Artificial Intelligence and Computing (ICAAIC 2023)

IEEE Xplore Part Number: CFP23BC3-ART; ISBN: 978-1-6654-5630-2

Traffic Light Detection System in Self-Driving Cars

P. Kanchanamala A. Dharani J. Siri Chandana
Information Technology Information Technology Information Technology
GMR Institute Of Technology GMR Institute Of Technology GMR Institute Of Technology
Rajam, Andhra Pradesh, India Rajam, Andhra Pradesh, India Rajam, Andhra Pradesh, India
kanchanamala.p@gmrit.edu.in 19341a1204@gmrit.edu.in 19341a1249@gmrit.edu.in
2023 2nd International Conference on Applied Artificial Intelligence and Computing (ICAAIC) | 978-1-6654-5630-2/23/$31.00 ©2023 IEEE | DOI: 10.1109/ICAAIC56838.2023.10140631

D. Jyothsna I. Sudeep Kumar G. Kushila

Information Technology Information Technology Information Technology
GMR Institute Of Technology GMR Institute Of Technology GMR Institute Of Technology
Rajam, Andhra Pradesh, India Rajam, Andhra Pradesh, India Rajam, Andhra Pradesh
19341a1235@gmrit.edu.in 19341a1248@gmrit.edu.in 19341a1246@gmrit.edu.in

Abstract: In recent years, Self-dri vi ng cars have been a important features fro m the input image and helps in
topic of interest. These self-dri vi ng cars can opti mize removing the false positives [4]. For the classification
their dri ving behavior to maxi mize fuel efficiency and purposes SVM can be used which performs binary
reduce traffic congestion. The paper discusses the classification and gives more accuracy by div iding inpu t
challenges of traffic light detection in self-dri vi ng based on day and night conditions [5]. Faster RCNN saves
vehicles and suggests ways to address them. Traffic light the time when compared to tradit ional algorith ms and gives
detection is a critical component of self-dri vi ng car the more accuracy [6]. It is also impo rtant to measure the
technolog y, as it enables autonomous vehicles to computation time, by using the Haar-like features the
accurately percei ve and res pond to traffic signals. An calculation speed will get increased and able to detect traffic
overview of traffic light detection is provi ded in self- lights in less time [7]. SVM creates a hyperplane and d ivide
dri vi ng cars, outlining its benefits and drawbacks. It input image into two classes based on day and night
reviews the existing literature on traffic light detection conditions and hance can able to detect traffic lights at night
and highlights vari ous approaches, includi ng computer times also with mo re accuracy [8]. The sensors and cameras
vision methods, machine learning techni ques, and deep capture images of the surrounding environment and process
learning models that have been proposed to address this the information to identify the location of t raffic lights in the
issue. The algorithms like SVM, Heuristic and CNN scene [9].
algorithms are proposed i n the existing soluti ons. These According to earlier studies, [10] a self-driving car is
algorithms work accurate during day ti me but not i n capable of detecting traffic lights within a range of 13 to 115
night time. And also, by using many di fferent sensors for meters. In conclusion, traffic light detection is an essential
detection the main drawback is that they can fail over technology for self-d riv ing cars, allowing them to safely and
time which may lead to costly replacements. This study effectively navigate roadways and crossings. With further
examined vari ous traffic light detecti ng algorithms and advances in machine learning and co mputer vision, traffic
their procedures. Also note the difficulties and upcoming light detection technology will continue to improve, making
work on traffic signal detecti on for precise findings. autonomous driving systems even more reliable and
SVM is frequently empl oyed, and CNN and Faster effective.
RCNN are well-known algorithms for object detection
and traffic signal color categorization. II. RELATED WORK
Diaz-Cabrera et al. developed a method for detecting traffic
Keywords: Convolutional Neural Network, Lisa Dataset,
lights and estimat ing their distance in both day and night
Traffic lights, image Processing, Support vector Machine.
conditions. The method combines colo r analysis, lu minosity
I. INT RODUCT ION adaptation, fuzzy clustering, aspect ratio analysis, tracking,
Autonomous driving technologies that allow self-driving and Bayesian filters to achieve 99.4% accuracy in detecting
cars to safely negotiate roads and intersections depend on traffic lights fro m 10-115m. The system was tested in real-
traffic light detection. [1]. Traffic lights are critical traffic world urban scenarios for over an hour and validated during
control devices that provide vital info rmation to drivers and a public road urban driverless car test in Italy in 2013. The
pedestrians, indicating when to stop or go. Therefore, authors' main contribution is a new algorithm that detects
detecting traffic lights accurately and in real-time is crucial
traffic lights using a combination of adaptive control, fuzzy
for the safe and efficient operation of self-driving cars [2].
There are several algorithms p resent for detecting the traffic clustering, and sequential rules for colo r segmentation, and
lights in self-d iving cars. An algorithm called Histogram o f Bayesian filtering for distance estimation. The algorith m is
Gradient Features is used to extract sharp edge features reproducible and has reduced computational times due to
effectively and can ab le to work with h igh-dimensional data parallelization.[1]
[3]. Convolutional Neural Network helps in extracting the

978-1-6654-5630-2/23/$31.00 ©2023 IEEE 885

Authorized licensed use limited to: SRM University Amaravathi. Downloaded on July 24,2023 at 05:48:51 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the Second International Conference on Applied Artificial Intelligence and Computing (ICAAIC 2023)
IEEE Xplore Part Number: CFP23BC3-ART; ISBN: 978-1-6654-5630-2

Kim et al. presented a method to identify traffic lights at Advanced driver assistance systems and driverless vehicles
night by extracting candidate blobs from an RGB colo r may be able to recognize traffic signals better using this
image, representing the image on the dominant color technique. [5]
domain, selecting red and green color reg ions as candidates, Based on a recently annotated collection of traffic scene
filtering the candidates by using shape information, and image data, Chen et al. described a method for detecting and
classifying the candidates using a SVM classifier with basic identifying traffic signals using Faster RCNN. Through
features, bright-based stochastic features, and geometric - experimental co mparison, the best feature extraction
based mo ment's values. The system was imp lemented on an network was chosen, and the method's efficacy was tested
Intel Core CPU and tested on urban and rural road v ideos, on a platform for self-driv ing cars. The Faster RCNN
showing a 93% detection rate with an average co mputation algorith m, wh ich is used in this model, has four steps:
time of 15ms/frame. In this model the classificat ion done by candidate region generation, feature ext raction,
multi-classifier by SVM algorithm.[2] classification, and position correction. It is more accurate
Xinyu et al. utilized a pre -installed camera that was placed and quicker than the Fast RNN algorith m. A deep network
on the inside of a vehicle's windshield to identify traffic framework unifies these actions. The first phase creates
lights. They trained a computer system to recognize potential regions, while the second step uses the input traffic
different types of traffic lights by taking appro ximately scene image to extract features to produce regional scores.
5,000 images of 13 different categories of traffic lights and In order to get region proposal features, the scores are then
transforming them into high-dimensional features using a prioritized using a non-maximu m suppression approach and
"HOG descriptor." They then used these samples to train a input into the ROI pooling layer. Finally, a fully linked layer
hierarchical SVM classifier. The algorithm consists of three that outputs the categorizat ion score and area location using
main stages: pre-processing, detection, and recognition. Pre - these features as input. As a result, this model analyses the
processing involves preparing the input data for the performance of three various feature extract ion networks,
algorith m, detection involves identifying traffic lights in an choosing the best one through testing. Real-world tests on
image, and recognition involves identifying the specific type an intelligent driving platform were used to validate the
of traffic light that has been detected. [3] suggested strategy. [6]
To provide real-t ime traffic signal identification fo r Jung et al. discussed a technique for detecting traffic lights
autonomous vehicles, Ouyang et al. integrated a lightweight during the day. It involves three different steps: first,
CNN classifier model with a heuristic ROI detector. In Potential Traffic Lights Detector (PTLD) looks at an
addition to two autonomous cars, they also set up the model image’s color channel to find green and red traffic lights in
on a portable platform for testing on the road. Second, based the image. Second, the Shape Filter (SF) removes noise like
on current deep learning advancements, they created a CNN other signs, trees, vehicles, and build ings. It looks at the size
architecture for the categorization tasks. Using two publicly and shape of the different blobs it detects. Finally, Adaptive
accessible traffic light datasets as well as their proprietary Multi-class Classifier (AM C) classifies the detected traffic
dataset, this approach can outperform existing algorith ms. lights using Haar-like features and the Adaboost algorithm
As a conclusion, they uncover issues with the present following an intermediary association phase that identifies
publicly available dataset, such as biased sampling and data pertinent locations. [7]
imbalance. They provide a variety of strategies for d ata Ozcelik et al. first taken images using CCD camera. Then
collection and pre-processing as a resolution. [4] they applied Median Filter algorith m. The RGB image is
Sani et al. developed a method for detecting and recognizing converted to HSV format, fo llo wed by a co lor-based
traffic signal arrow symbols during the day using digital segmentation method and morphological procedures, in
image processing and machine learning. They used the order to identify the traffic lights in the incoming image.
HOG descriptor fo r feature extraction and the Support Color segmentation is a process used to separate different
Vector Machine classifier for learn ing and classification. objects in an image based on their color while
The proposed approach achieved a high accuracy of morphological operations are used to refine the image and
98.52%, precision of 98.52%, and recall of 98.68%. The apply certain effects to it. The process of edge extraction is
method co mbines machine learn ing and image processing to used to highlight the edges of objects and the Hough Circle
distinguish different types of traffic signal arrows using the algorith m is used to detect circular objects in the image.
SVM classifier. Each kind of t raffic signal arrow in the After co mpleting all of these steps, the traffic light in the
training set underwent pre-processing, detection, and the image's state is ascertained using a machine learn ing method
extraction of HOG descriptors before being used to train the like Support Vector Machines (SVM). [8]
SVM classifier. The perfo rmance of the trained SVM An innovative, lightweight deep CNN model, known as
classifier was assessed using the test data, and the method MBBNet, was put forth by Ouyang et al. for traffic light
was then used to process a recorded input video to create an recognition in self-driv ing cars. MBBNet uses channel
output video showing the detected traffic signal arrows. compression to manage model parameters while retaining

978-1-6654-5630-2/23/$31.00 ©2023 IEEE 886

Authorized licensed use limited to: SRM University Amaravathi. Downloaded on July 24,2023 at 05:48:51 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the Second International Conference on Applied Artificial Intelligence and Computing (ICAAIC 2023)
IEEE Xplore Part Number: CFP23BC3-ART; ISBN: 978-1-6654-5630-2

accuracy and combines many cutting-edge convolutional and recognize traffic lights. The system runs through three
backbones. The model was tested using data from actual phases:
road conditions, and it demonstrated good accuracy for self - 1. Candidate region extract ion: Color threshold
driving buses (accuracy > 0.94 and Average IOU > segmentation is used to extract the candidate regions of
74.05%). The model in this paper uses a two-step traffic traffic lights from the camera images in this stage.
light detection methodology, with MBBNet acting as the 2. Noise removal and filtering: Two methods of shape-based
detector in the first stage and a straightforward CNN filtering are done to the candidate regions to reduce noise
classifier based on color space fusion serving as the detector and increase detection accuracy.
in the second stage to identify the different types of traffic 3. Template matching: The candidate regions are then
lights. MBBNet uses channel compression to create a verified as traffic lights using template matching with
smaller, quicker model and integrates mu ltip le normalized cross correlation approaches.
convolutional backbones for greater accuracy. Fo r effective The experimental findings show that the suggested method
categorization, the classifier uses a fused input of RGB and is reliable and efficient for traffic signal detection in
HSV color spaces. When employing huge raw input photos, intricate urban settings.[13]
autonomous cars can operate more efficiently overall by Zhang Yue et al. detected potential targets for traffic signals
cutting the first stage's processing time. [9] inside a zone of interest defined by the vanishing line by
By using learning data to identify the candidate area in the combin ing a nu mber of variab les, including colo r
traffic light image, Lee et al. developed Haar-like features. segmentation, blob recognition, and structural feature
The SVM (Support Vector Machine) classifier uses prio r extraction. A fusion technique is used to combine the
learning to validate the discovered candidate image. The findings fro m each of these features and award a score to
next step is binarization processing, a two-class each prospective location indicating the likelihood that it
classification that establishes a cutoff point to separate the will include a traffic signal. Further integrating the score of
image's pixels into black and white. Then, to eliminate each candidate place with temporal and geographic data
extraneous noise and accentuate the traffic light, the results in the final detection. As a result, the algorith m is
morphological processes of erosion and dilat ion are applied. guaranteed to be effective and efficient at detecting traffic
To ascertain whether traffic lights are now on or off, the signals in actual situations. The suggested method has
observed traffic light is separated into corresponding signal undergone rigorous testing on an actual intelligent car, and
zones. [10] the results have demonstrated that it is successful in
Huang et al. suggested a technique for detecting traffic identifying traffic lights. [14]
lights that uses color extraction, template matching, and a A method of image p rocessing was developed by
support vector machine. Colo r ext raction is used to locate Nashashibi et al. to find and identify traffic lights. Their
potential candidates for traffic lights. These candidates are approach makes use of grayscale spot light detection to
then checked for size consistency and appropriate asp ect detect traffic lights and adaptive templates to identify them,
ratio. Template matching is emp loyed to detect traffic lights enabling the identificat ion of distinct traffic signal kinds
by comparing an area with a reference area of that light fro m various nations. Because of the system's modular
being off. The Support Vector Machine is used to find design, it may be versatile in recognizing various traffic
optimal settings from the collected data, which leads to signal kinds. A number of traffic light recognition systems
more accurate detection results.[11] were also created by them utilizing object recognition
The model was divided into two parts by Guidolini et al., techniques like cascade classifiers and AdaBoost for
which included offline map generation and annotation as comparison. Using a prototype car and video clips fro m
well as online detection and recognition of traffic signals. several nations, the system was evaluated in actual driving
IARA (Intelligent Autonomous Robotic Automobile), the situations. The results showed a high accuracy of correctly
vehicle being used to test the suggested method, is outfitted detected traffic signals and a low amount of false alerts. By
with a number of sensors, including cameras, LiDA R, and offering precise and trustworthy real-time informat ion about
IMU. Using colored photos as input, the deep learning traffic signals, the suggested system can increase the safety
model YOLOv3 is utilized to identify the traffic lights and of intelligent cars and driver support technologies. Because
categorize their condition. The earlier maps are made up of of the system's ability to detect various traffic signal types
informat ion gathered while IA RA is manually d riven along fro m different nations, it may also be adjusted to different
the desired route and saved in a Road Definit ion Data File geographical areas. Using 640x480 images, the processing
(RDDF). The DBSCAN technique is then used to filter and was done in real-time on a desktop computer running at
cluster these map data together with the LiDAR points. [12] 2.9GHz. [15]
The technology described by Wang et al. seeks to offer an Liu et al. developed a new method for detecting traffic lights
effective and timely technique for autonomous cars to detect with high accuracy, even in challenging lighting conditions.
Their technique co mbines computer v ision and machine

978-1-6654-5630-2/23/$31.00 ©2023 IEEE 887

Authorized licensed use limited to: SRM University Amaravathi. Downloaded on July 24,2023 at 05:48:51 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the Second International Conference on Applied Artificial Intelligence and Computing (ICAAIC 2023)
IEEE Xplore Part Number: CFP23BC3-ART; ISBN: 978-1-6654-5630-2

learning, leveraging GPS to locate the traffic light's region

of interest. They emp loy a convolutional neural network to
analyze camera images and produce a saliency map, which
highlights the traffic light's location and improves
recognition performance, especially in low-light scenarios.
The algorith m's effect iveness was tested against a reference
algorith m, using real-world datasets in varied conditions.
The results demonstrate that the proposed method performs
well under different lighting conditions and is suitable for
advanced driver assistance systems and autonomous vehicle
navigation. This innovation addresses the critical need fo r Fig.1. CNN's T raffic Light Detection T echnology for Autonomous
precise traffic light detection, wh ich is essential fo r Vehicles
autonomous vehicle safety and real-t ime info rmation B. Support Vector Machine:
delivery to drivers via advanced driver assistance systems.
Support Vector Machine is an algorithm which is mainly
[16] used for classification tasks. This method is efficient fo r
A trustworthy model fo r traffic signal detection on on -
detecting the traffic lights in night t ime. The algorithms
vehicle cameras is p resented by Huimin Ma et al. with the work by creating a hyperplane that separate the data into
aim of enhancing computational efficiency and recognition different classes. And by using these features the hyperplane
precision. In addit ion to introducing several imp ro ved
predict the class of new images by transforming the images
techniques built on the aggregate channel feature approach, into feature maps. SVM detects the traffic light through
they propose a task model that takes into account the aspect following steps.
ratio, area, location, and context of traffic lights. These
a) Selecting the Candidate Regions: Fro m an RGB
include developing a network of fusion detectors and
color image, the candidate regions of the traffic light are
customizing features for various types of traffic lights. The retrieved. The image is represented on a dominant color
authors also present an innovative inter-frame info rmation domain, and the red and green colors are extracted via a
analysis method that enhances precision by using detection labelling technique.
data from the preceding frame. The authors exhib it b) Noise Reduction Step: This phase emp loys shape
competitive results by contrasting their traffic light detection filtering to eliminate the noisy blobs. To accurately detect
model to others using the VIVA dataset. This approach has traffic signals, the areas that contain other types of lights,
the potential to increase the security of on-vehicle cameras such as streetlights, traffic signs, and vehicle lamps, are
removed.
by precisely recognizing traffic signals and giving drivers
c) Multi-class Classifier: It separates the traffic light
real-time information. [17] blobs from candidate blobs. The hyper plane is created
which performs b inary classification and classify the data
into two classes based on the features on input data.
III. VARIOUS METHODS OF TRAFFIC LIGHT
DETECTION SYSTEM C. Faster R-CNN:
A. Convolutional Neural Network (CNN): The Faster R-CNN is trained on traffic light images and
helps in detecting and classifying the traffic lights in new
Convolutional Neural Net work is a deep Learn ing technique
mainly used in image classificat ion tasks. Here the input images. This algorithm works in two stages:
image is convolved with mult iple filters to ext ract the a) Region Proposal Network: A set of candidate
various filters and patterns. These feature maps are passed regions is created by the RPN that contains the objects of
interest. The RPN is a light-weight CNN that slides a small
through several fully connected layers to make a p rediction
network over the feature maps of the input images to predict
about the class label. The final layer of CNN outputs a object boundaries.
probability distribution over the position of class labels and b) Classification and Refinement: The second stage
the class label with highest probability is selected as the uses a full CNN to classify the candidate regions generated
final prediction. by the RPN and refine their bounding box locations. This
Steps involved in CNN are: stage takes the features extracted fro m candidate regions and
a) Convolution layer: Apply filters to input image to used them to make prediction about the class of each region .
extract features.
b) Pooling layer: Squeezing the feature maps spatial
dimensions.
c) Fully Connected layer: Predicting the class label
using the retrieved characteristics.

978-1-6654-5630-2/23/$31.00 ©2023 IEEE 888

Authorized licensed use limited to: SRM University Amaravathi. Downloaded on July 24,2023 at 05:48:51 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the Second International Conference on Applied Artificial Intelligence and Computing (ICAAIC 2023)
IEEE Xplore Part Number: CFP23BC3-ART; ISBN: 978-1-6654-5630-2

In the detection phase, Adaboost applies the strong

classifier to an image and outputs a class label for each
Region Of Interest (ROI), indicating whether it is a traffic
light or not. The algorith m also provides a confidence score
for each classification, which can be used to filter out false
positive detections. Overall, Adaboost is a highly effective
technique for traffic light detection in self-driving cars, as it
combines mult iple weak classifiers into a strong classifier
that can detect traffic lights with high accuracy in real-world
conditions.

Fig.2. Flow chart of Faster RCNN in T raffic Light Detection IV. CHALLENGES
There are Several algorith ms available for detecting the
D. YOLO v3:
traffic lights in self-driving cars among those while
It works by div iding an image into a grid and then making detecting the traffic lights several challenges are faced.
predictions for each grid cell about the presence of an object CNN algorith m which is based on image processing have
and its location. The YOLO model is trained to recognize some difficulty in detecting traffic lights like co mplex
the shape and color of t raffic lights and make pred ictions environment conditions like lighting conditions, reflections
based on the presence of these features in a grid cell. Du ring which may affect the accuracy of traffic light detection [1].
inference, the model outputs the bounding boxes and class High Dimensional input data which is a challenging task fo r
probabilit ies after receiv ing an image as input for any SVM algorithm [2]. Labelling and annotating a large dataset
detected traffic lights. The algorith m uses a single for training YOLOv3 can be t ime-consuming and
convolutional neural network to make predict ions for the challenging which can limit the overall performance o f the
entire image. model [12]. Overfitting can occur when training an SVM on
E. Histogram of Gradient Features: limited training data, which can lead to poor generalization
performance on new and unseen data [11]. The different
An image of a traffic light would be used to extract features shapes of traffic lights and size of the traffic lights vary
using HOG, and then uses classifier such as SVM fo r fro m p lace to place becomes a difficu lty in detecting traffic
classifying traffic lights. It computes the gradient orientation lights [10]. In some cases, CNN may produce the false
histograms for each cell after divid ing an image into tiny detections, such as detecting a street sign or a reflection as a
cells. The g radient orientation histograms represent the traffic light, leading to incorrect decisions by self-driving
distribution of edge directions in each cell, and they capture cars [13]. Haar-like features can work only in gray-scale
images and lack the ability to describe colors [10]. Adaboost
the local texture and shape information of the image.
algorith m takes more t ime for feature extraction which is
These HOG features are then concatenated into a feature time consuming and challenging for t raffic light detection
vector, which can be used as input to a classifier. The and also there are chances of overfitting when training a
classifier takes the HOG features as input and outputs a model leads to the poor performance in detection [10].
prediction about the class of the object in the image (red, Faster RCNN rely on deep neural networks to extract
yellow, or green). features, which can be sensitive to variations in image
quality and lead to poor performance [6]. HOG which is
sensitive to image rotations these are not inherently scale
and rotation invariants, which means they can perform
poorly when detecting traffic lights that are not at a fixed
scale or o rientation in an image [3]. R-CNN requires large
amount of training data and takes more time to train the
model making them challenging to use in applications where
real-time performance is a requirement [3].

Fig.3. Object Classification Using Histogram of Gradient Features

F. Adaboost Algorithm:
Adaboost is used to combine mu ltip le weak classifiers into a
strong classifier that can accurately detect traffic lights in
real-t ime. The algorith m works by training the weak
classifiers on positive and negative examples of traffic
lights, and then assigning a weight to each classifier based
on its accuracy. The classifiers with higher accuracy receive
a higher weight, and the final strong classifier is created by
combining the weighted output of the individual classifiers.

978-1-6654-5630-2/23/$31.00 ©2023 IEEE 889

Authorized licensed use limited to: SRM University Amaravathi. Downloaded on July 24,2023 at 05:48:51 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the Second International Conference on Applied Artificial Intelligence and Computing (ICAAIC 2023)
IEEE Xplore Part Number: CFP23BC3-ART; ISBN: 978-1-6654-5630-2

V. CONCLUSION [13] Wang, C., Jin, T., Yang, M., & Wang, B. (2011). Robust and real-time
traffic lights recognition in complex urban environments. International
The traffic signal detection system is a crucial applicat ion of Journal of Computational Intelligence Systems, 4(6), 1383-1390.
computer vision that has the potential to significantly [14] Zhang, Y., Xue, J., Zhang, G., Zhang, Y., & Zheng, N. (2014, July). A
improve traffic management and road safety. Modern image multi-feature fusion based traffic light recognition algorithm for intelligent
vehicles. In Proceedings of the 33rd Chinese control conference (pp. 4924 -
processing methods and mach ine learning algorithms are 4929). IEEE.
used to accurately recognize and categorize traffic signals in [15] De Charette, R., & Nashashibi, F. (2009, October). Traffic light
real-t ime. These technologies are made to be ab le to identify recognition using image processing compared to learning processes. In
the position and status of traffic lights. A number o f 2009 IEEE/RSJ International Conference on Intelligent Robots and
techniques, including SVM, HOG, CNN, and Ada-boost Systems (pp. 333-338). IEEE.
[16] John, V., Yoneda, K., Qi, B., Liu, Z., & Mita, S. (2014, October).
algorith ms, are used to identify traffic signals. Each Traffic light recognition in varying illumination using deep learning and
algorith m has unique benefits and drawbacks. A lthough saliency map. In 17th International IEEE Conference on Intelligent
these algorith ms are capable of producing precise detections T ransportation Systems (ITSC) (pp. 2286-2291). IEEE.
throughout the day, they perform worse when the light is [17] Li, X., Ma, H., Wang, X., & Zhang, X. (2017). Traffic light
off. In order to overcome th is challenge, future study will recognition for complex scene with fusion detections. IEEE Transactions
on Intelligent Transportation Systems, 19(1), 199-208.
suggest a model that groups the photos according to whether
it is day or night using CNN and fuzzy clustering. Also, this
helps the model's accuracy when applied at night.

VI. REFERENCES
[1] Diaz-Cabrera, M., Cerri, P., & Medici, P. (2020). Robust real-time
traffic light recognition and distance estimation using a single camera.
Expert Systems with Applications, 42(8), 3911-3923.
[2] Kim, H. K., Shin, Y. N., Kuk, S. G., Park, J. H., & Jung, H. Y. (2013).
Night-time traffic light detection based on svm with geometric moment
features. International Journal of Computer and Information
Engineering, 7(4), 472-475.
[3] Mu, G., Xinyu, Z., Deyi, L., T ianlei, Z., & Lifeng, A. (2015). Traffic
light detection and recognition for autonomous vehicles. The Journal of
China Universities of Posts and T elecommunications, 22(1), 50-56.
[4] Ouyang, Z., Niu, J., Liu, Y., & Guizani, M. (2019). Deep CNN-based
real-time traffic light detector for self-driving vehicles. IEEE transactions
on Mobile Computing, 19(2), 300-313.
[5] Sani, Z. M., & Isa, M. S. H. (2020). Traffic Light Arrow Shape
Recognition Using HOG Descriptors and SVM Classifiers. International
Journal of Electrical Engineering and Applied Sciences (IJEEAS), 3(2), 53-
60.
[6] Pan, W., Chen, Y., & Liu, B. (2019, December). Traffic Light Detection
for Self-Driving Vehicles Based on Deep Learning. In 2019 15th
International Conference on Computational Intelligence and Security (CIS)
(pp. 63-67). IEEE.
[7] Kim, H. K., Park, J. H., & Jung, H. Y. (2011). Effective traffic lights
recognition method for real time driving assistance systemin the
daytime. International Journal of Electrical and Computer
Engineering, 5(11), 1429-1432.
[8] Ozcelik, Z., Tastimur, C., Karakose, M., & Akin, E. (2017, October). A
vision-based traffic light detection and recognition approach for intelligent
vehicles. In 2017 International of Computer Science and Engineering
(UBMK) (pp. 424-429). IEEE.
[9] Ouyang, Z., Niu, J., Ren, T., Li, Y., Cui, J., & Wu, J. (2020). MBBNet:
An edge IoT computing-based traffic light detection solution for
autonomous bus. Journal of Systems Architecture, 109, 101835.
[10] Lee, S. H., Kim, J. H., Lim, Y. J., & Lim, J. (2018, January). Traffic
light detection and recognition based on Haar-like features. In 2018
International Conference on Electronics, Information, and Communication
(ICEIC) (pp. 1-4). IEEE.
[11] Chen, Z., Shi, Q., & Huang, X. (2015, June). Automatic detection of
traffic lights using a support vector machine. In 2015 IEEE Intelligent
Vehicles Symposium (IV) (pp. 37-40). IEEE.
[12] Possatti, L. C., Guidolini, R., Cardoso, V. B., Berriel, R. F., Paixão, T.
M., Badue, C., ... & Oliveira-Santos, T. (2019, July). Traffic light
recognition using deep learning and prior maps for autonomous cars. In
2019 international joint conference on neural networks (IJCNN) (pp. 1-8).
IEEE.

978-1-6654-5630-2/23/$31.00 ©2023 IEEE 890

Authorized licensed use limited to: SRM University Amaravathi. Downloaded on July 24,2023 at 05:48:51 UTC from IEEE Xplore. Restrictions apply.