Parking Occupancy Detection using Thermal Camera

Vijay Paidi and Hasan Fleyeh

Dalarna University, School of Technology and Business Studies, Borlänge, Sweden

Keywords: Convolutional Neural Network, Detectors, Thermal Camera.

Abstract: Parking a vehicle is a daunting task during peak hours. The search for a parking space leads to congestion and

increased air pollution. Information of a vacant parking space would facilitate to reduce congestion and

subsequent air pollution. This paper aims to identify parking occupancy in an open parking lot which consists

of free parking spaces using a thermal camera. A thermal camera is capable of detecting vehicles in any

weather and light conditions based on emitted heat and it can also be installed in public places with less

restrictions. However, a thermal camera is expensive compared to a colour camera. A thermal camera can

detect vehicles based on the emitted heat without any illumination. Vehicles appear bright or dark based on

heat emitted by the vehicles. In order to identify vehicles, pre-trained vehicle detection algorithms, Histogram

of Oriented Gradient detectors, Faster Regional Convolutional Neural Network (FRCNN) and modified Faster

RCNN deep learning networks were implemented in this paper. The detection rates of the detectors reduced

with diminishing of heat in the vehicles. Modified Faster RCNN deep learning network produced better

detection results compared to other detectors. However, the detection rates can further be improved with larger

and diverse training dataset.

1 INTRODUCTION

Parking can be a daunting task during peak hours due

to availability of limited number of parking spaces

compared to the number of vehicles. It can lead to

congestion and increased air pollution. Parking

management is an integral component for city

planning administrators and this is one of the research

themes in a smart city development. It can take up to

14 minutes to find a parking space according to

previous studies (Shoup, 2006, Polycarpou et al.,

2013). Parking has been an important research area as

it enables accessibility to commuters and is capable

of enhancing business opportunities (Paidi et al.,

2018). Parking problem is applicable to both major

and minor cities due to higher demand and limited

resources, though there can be a difference in demand

between different places. Lack of parking spaces at a

retail store can lead to loss of business opportunities.

Therefore, stores spend higher expenditure to acquire

sufficient number of parking spaces which cover

large areas of scarce land resources. Due to space

restrictions, in urban areas, new parking areas are

being built in multi storey buildings or basements.

Another way to address parking problem is to use

public transport which might not be a convenient

option for everyone.

Parking guidance systems were developed to

reduce congestion, fuel costs and air pollution. There

are several web and mobile smart parking

applications available. Smart parking applications are

available for closed parking lots while there are no

applications available for open parking lots (Paidi et

al., 2018). Demand of parking spaces in an open

parking lot is higher compared to a closed parking lot.

Therefore, this paper aims to detect parking

occupancy in an open parking lot. An open parking

lot is also subjected to different environment

conditions such as; snow, rain, darkness and sunny.

An investment made in parking guidance system or

smart parking application for open parking lots

cannot be retained directly as the parking spaces are

available freely. There are several tools used to

identify parking occupancy such as; infrared and

ultrasonic sensors, magnetometers, vehicle ad hoc

networks (VANET) and microwave radars (Hassoune

et al., 2016, Idris et al., 2009). However, sensors,

VANET, microwave radars and magnetometers need

higher expenditure in installation and maintenance

activities. Therefore, usage of camera is a suitable

technology to identify parking occupancy which is

capable of covering large number of parking spaces

with a single camera. However, the number of

parking spaces covered by a camera is also dependent

Paidi, V. and Fleyeh, H.

Parking Occupancy Detection using Thermal Camera.

DOI: 10.5220/0007726804830490

In Proceedings of the 5th International Conference on Vehicle Technology and Intelligent Transport Systems (VEHITS 2019), pages 483-490

ISBN: 978-989-758-374-2

483

on the height and angle of the positioned camera. In

this paper, a thermal camera is used for obtain parking

occupancy detection in an open parking lot. An open

parking lot is subjected to various weather conditions

and a thermal camera is capable of detecting objects

during such conditions. A thermal camera is

expensive compared to a colour camera but it has less

privacy restrictions compared to the use of normal

colour camera and it is suitable for any environment

and light conditions. There is not much previous

literature available on the use of thermal cameras to

identify parking occupancy. Therefore, this paper

aims to address this research gap with this paper. Pre-

trained detectors, trained detectors using aggregate

channel features, histogram of oriented gradient and

deep learning network are implemented for parking

occupancy detection. The remaining sections of the

paper is organized in the following way. Section 2

discusses relevant literature while Section 3 discusses

about the dataset and set of detectors used in this

paper. Section 4 presents the results obtained in this

paper along with analysis. Section 5 discusses pros

and cons of using a thermal camera for parking

occupancy detection. Finally the paper is ended with

conclusion in Section 6.

2 LITERATURE REVIEW

Since, parking occupancy detection is performed

using a camera, relevant algorithms or detectors are

discussed in this section. The use of a thermal camera

for identifying parking occupancy is not discussed in

previous literature. However, when it comes to colour

camera, there are several studies available in the

literature. Edge detectors such as Canny and Sobel

can be used to identify vehicles (Bao et al., 2005)

(Maini and Aggarwal, 2009). Edge detectors perform

efficiently when the vehicles can be recognized.

However, when using thermal cameras, due to loss of

heat in the vehicle over a period of time, vehicles can

become dark and can be challenging to recognize or

detect. In such scenarios edge detectors might not be

a suitable option for vehicle detection. Histograms of

Oriented Gradient (HOG) descriptors and Viola Jones

are efficient human or pedestrian detectors (Dalal and

Triggs, 2005) which can also be used for vehicle

detection. HOG detector trains using positive and

negative images using a linear Support Vector

Machine (SVM) classifier (Mao et al., 2010). HOG

detector is invariant to geometric and photometric

transformations (Xu et al., 2016) and is therefore, one

of the detector tested in this paper.

Neural networks is an evolving data processing

system used for classification purposes. It is inspired

by the human brain nervous system. In one study,

neural networks are used for efficient recognition of

license plates (Villegas et al., 2009). Convolutional

neural networks is a deep learning method for

classification of images which can be used for object

detection and it can also be adapted for identifying

parking occupancy (Amato et al., 2017). Deep

learning is capable of handling complex object

detections and is therefore one of the method tested

in this paper. Increased performance normally

requires higher computational costs and in order to

maintain accuracy with lower computational costs a

binary sliding window detector also called aggregated

channel feature (ACF) detector was developed

(Dollár et al., 2014, Liu and Mattyus, 2015). An

image is computed to multiple channels and then sum

every block of pixels to generate lower resolution

channels. A multi scale sliding window is employed

and boosting is performed to identify object in ACF.

Kalman filter is used to identify and classify night

time traffic surveillance (Robert, 2009). Headlight

and visible vehicle features are used to detect

vehicles. Similarly in another study, (Fleyeh and

Mohammed, 2012) vehicles during night time were

detected using blob properties which were classified

by an SVM classifier and Kalman filters were

invoked to track the detected vehicles. In another

study, infrared thermal camera was used to study

traffic flow by using Viola jones detector. The tires

and windshield of the vehicles were used to identify

the vehicles (Iwasaki et al., 2013). Based on the

chosen features, positive and negative images were

used to train the detector. However, since the vehicles

are moving on the road, heat can be captured by the

windshield or by the tires which might not be

applicable to this paper since the vehicles are

stationary in a parking lot. In a similar study, an

unmanned aerial vehicle with help of a thermal

camera is deployed for detection of people and cars

(Gaszczak et al., 2011). Detection of people and cars

are performed using multiple cascaded Haar

classifiers. Haar uses a set of weak classifiers to form

a strong classifier. However, the position and angle of

the vehicle affected the detection using this classifier.

3 METHOD

The data for parking occupancy detection for this

paper is captured using a thermal camera. Thermal

camera is normally deployed for surveillance

purposes where objects are identified using heat

VEHITS 2019 - 5th International Conference on Vehicle Technology and Intelligent Transport Systems

484

Figure 1: Images of parking space from thermal camera.

signatures (Robert, 2009). The thermal camera

installed in this paper is equipped with 19mm focal

length objective on top of a two storey building. Since

the camera was not placed at sufficient height the

entire parking lot was not covered. Therefore, due to

height and focal length limitations of the camera, only

first four rows of the parking lot were selected to

identify parking occupancy. The data was collected in

different weather conditions such as snow, rain, dark

and bright conditions. In Scandinavian countries or

other similar countries, daylight is shorter during

winter and longer during summer. Snow can be seen

for three to four months during winter. A colour

camera can face issues with low light levels and snow

conditions which are common in Scandinavian

countries. Based on the privacy policy guidelines in

Sweden, it is restricted to use video surveillance in

open public areas where individuals can be

recognized or identified (Datainspektionen, 2018).

Therefore, the use of thermal camera had helped to

avoid these restrictions where no individuals can be

identified or recognized. There are multiple parking

spaces in the parking lot selected for this study.

However, the first four rows of the parking lot is

identified as the region of interest which is marked by

the green rectangle as shown in Figure 2. The vehicles

outside region of interest are small and majorly

occluded which was the reason for exclusion. The

thermal camera records videos based on motion

detection which lead to collection of several small

interval videos. The videos were collected in bright,

dark and snowy weather conditions. Videos from

various weather conditions were selected and one

frame from each video was collected. In this way 60

images were collected from 60 videos.

Figure 2: Area of parking spaces.

In the morning, vehicles are warm and can be

easily identified during winter or sunny conditions.

However, when the heat in the vehicles reduce over a

period of time, they appear to be dark and are hard to

be recognized as shown in Figures 1.3 and 1.4. The

thermal camera uses pseudo colours to display data as

shown in Figures 1 and 2. Therefore, the data is

converted to grey scale for further processing. Since

there were no pre-labelled dataset available, occupied

and vacant parking spaces were labelled manually as

shown in Figure 3. Occupied spaces were labelled as

cars and unoccupied spaces were labelled as empty.

Diversity between the images was maintained to

improve feature detection. Due to limited number of

training images, all the images were used for training.

Testing was performed with a different set of images

which are not part of the training dataset. In order to

check the performance and accuracy of the detector,

images of different environmental conditions were

tested. There are free online databases such as PKLot

providing hundreds of parking lot images which can

be used to test performance of various algorithms (De

Almeida et al., 2015).

(1.1) Sunny in winter morning

(1.2) Dark snowy winter morning

(1.3) Dark winter evening

(1.4) Dark snowy winter afternoon

Parking Occupancy Detection using Thermal Camera

485

Table 1: Detectors description.

S.No

Detector

Description

Pre-trained Aggregate Channel Features

(ACF) detector

A trained vehicle detector based on aggregate channel

features. The detector is trained with unoccluded images of

vehicles (ACFdetector, 2018, Liu and Mattyus, 2015).

Pre-trained Faster

RegionalConvolutionalNeuralNetwork

(Faster RCNN) detector

A pre-trained deep learning convolutional neural

network trained using unoccluded images of vehicles. It

consists of modified version of CIFAR-10 architecture.

(FasterRCNN, 2018)

Trained ACF detector

A detector using ACF is trained using the dataset of 60

images captured by thermal camera. A new detector is

trained since, pre-trained detector is trained using colour

and unoccluded images. (ACFdetector, 2018)

Cascade detector

A cascade detector is trained using the dataset of 60

images. Features are trained using HOG which are efficient

in detecting objects compared to Haar or Binary Patterns.

(detector, 2018, Dalal and Triggs, 2005)

Faster RCNN

A Faster RCNN is trained using the thermal camera

dataset of 60 images. Since a pre-trained network is trained

using unoccluded images of vehicles, a new detector is

trained. The Faster RCNN is created using 11 layers. (Fan

et al., 2016). This detector is capable of producing better

results with less number of training images.

Faster RCNN from Resnet50

A Faster RCNN is created using a 50 layer deep

learning network (resnet50, 2018). Resnet is a trained deep

learning network which can classify 1000 objects. The

network is re-trained with the training dataset.

However, the images were collected by a colour

camera and the use of such images are not appropriate

for this paper. Therefore, only the images from the

thermal camera were used in training the detectors.

Figure 3: Labelled image.

4 RESULTS AND ANALYSIS

This section discusses and analyses the results

obtained from implementing the detectors.

4.1 Pre-trained ACF Detector

The pre-trained ACF detector was not able to detect

many vehicles as shown in Figure 4. Only one vehicle

was detected with a low score using the pre-trained

ACF detector. The vehicle in the middle of the

parking lot was only detected due to clear visibility.

Figure 4: Pre-trained ACF detector.

4.2 Pre-trained Faster RCNN Detector

The pre-trained Faster RCNN detector performed

better compared to ACF detector as shown in Figure

5. Approximately 33% of vehicles were detected

using this detector. Additional training needs to be

done to improve detection rate.

VEHITS 2019 - 5th International Conference on Vehicle Technology and Intelligent Transport Systems

486

Figure 5: Pre-trained Faster RCNN.

4.3 Trained ACF Detector

The ACF detector was trained using the images

available in the dataset. 60 images with two labels i.e.

cars and empty were used to train the ACF detector.

The ACF classifier was trained with 702 positive

images, 1404 negative images and 2048 weak

learners. However, due to limited number of training

images and noise, the detector was not able to detect

vehicles accurately. Large number of small areas with

high scores were detected which can be due to noise

as shown in Figure 6.

Figure 6:Trained ACF detector.

4.4 HOG based Cascade Detector

The cascade detector was trained with 671 positive

images and 1342 negative images. People, trees, etc.

were used as negative images for the training. Only

vehicle bounding boxes were selected for training.

However, as shown in the Figure 7, only one vehicle

is detected successfully while three vehicles were

partially detected. Therefore, detection rate of

vehicles with HOG detector is approximately 16%.

The position and angle of the parked vehicle is also

affecting the detection rate.

Figure 7: HOG based cascade detector.

4.5 Trained Faster RCNN Detector

The Faster RCNN deep learning network is trained

with the same training dataset. The network consists

of 11 layers and as shown in Figure 8, successful

detection rate is approximately 50%. Compared to the

previous algorithms, Faster RCNN achieved better

detection results. However, the network is newly

created and the available dataset is considered to be

small to achieve higher detection results. In order to

achieve better detection results, higher number of

images needs to be used for training.

Figure 8: Faster RCNN.

4.6 Modified Faster RCNN Detector

Since, large number of labelled images are not

available for training purposes, a Faster RCNN

network is updated from pre-trained deep learning

network. The pre-trained deep learning network is

Resnet-50 which consists of 50 layers and is trained

with thousands of images. Since the network is

already trained, it is experienced in extracting

features from images. Therefore, the last three layers

were modified and trained the network with available

dataset images. The results can be found in the three

images of Figure 9.

Parking Occupancy Detection using Thermal Camera

487

Figure 9(a): Detection results from sunny winter morning.

This detector performed better than the previous

detectors tested in this paper as shown in Figure 9(a).

The detection rate of vehicles achieved

approximately 88% while detection rate of empty

spaces is approximately 66%. Since the results were

considerable better, the deep learning network was

also tested on other environment conditions such as

winter afternoon and winter evening. In winter

afternoon, the cars became a bit cold and were not as

clear as in the morning. The detection rate of vehicles

is approximately 62% while detection rate of empty

spaces is approximately 33% as shown in Figure 9(b).

Figure 9(b): Detection results in winter afternoon.

In the evening, few vehicles were colder and appeared

darker leading to difficult detection results as shown

in Figure 9(c). The detection results of vehicles in

Figure 9(c) is approximately 42% while there no

successful detection of empty spaces. The modified

Faster RCNN performed better than other algorithms

tested in this paper. Pre-trained detectors did not

perform well as they were trained with unoccluded

colour images of vehicles. The ACF detector

aggregates blocks of pixels from various channels

such as; gradient histogram, gradient magnitude and

RGB channels.

Figure 9(c): Detection results from dark winter evening.

The images captured from the thermal camera are

based on the heat emitted from the object and does

not represent visual aspects of the image. Therefore,

aggregation of channel features method on pseudo

colour images or videos is not suitable for object

detection based on the Figure 4 and 6. The modified

Faster RCNN deep learning network performed better

in sunny or bright conditions where vehicles were

detectable. However, with diminishing of heat in

vehicles, the detection rate of vehicles and empty

spaces was reduced. The detection rate of vehicles in

bright conditions is nearly 88% while it is 42% in

dark conditions as shown in Figure 9(a) and 9(c). The

vehicles were largely occluded which also impacted

the detection rate. When the heat in the vehicles is

diminished, windshield, tires and headlights also

appear to be dark as shown in Figure 9(c). The

vehicles were detected in any environment if there is

heat in the vehicles as can be shown in Figure 9.

5 DISCUSSION

The training dataset used for the algorithms in this

paper consists of only 60 labelled images. Higher

number of training images can lead to better detection

results. Darker vehicles with less heat are also

detected using the modified Faster RCNN algorithm

as shown in Figure 9. However, only the first two

rows which consists of unoccluded vehicles were

detected successfully. The occluded vehicles in the

third and fourth row were not detected in dark winter

evening conditions where vehicles were colder and

darker. The pre-trained networks did not perform

efficiently on thermal images since they were trained

on colour images. The results clearly show that deep

learning network is capable of good detection results

in any environment if the vehicles are unoccluded.

The detection rate can be improved with occluded

VEHITS 2019 - 5th International Conference on Vehicle Technology and Intelligent Transport Systems

488

vehicles as well with higher number of training

images. Since the number of training images in this

paper is considered to be low, the detection results of

occluded vehicles and empty spaces was low. If the

position of the camera is placed at an increased height

or placed on top of the parking lot, unoccluded

images of vehicles can be obtained which can lead to

better detection rate in any environment or lighting

conditions.

A thermal camera can be used with less

restrictions compared to colour visual camera.

Privacy rules do not enable the free use of camera at

public places. The main drawback and advantage of

using a thermal camera is that the vehicles can be

recognized based on emitted heat. A vehicle can be

recognized during anytime or any environmental

conditions if there is sufficient amount of heat emitted

by the vehicle. This is one of the advantage of using

a thermal camera instead of a colour visual camera. If

a vehicle is stationary for a period of time, the heat in

the vehicle diminishes gradually and the vehicles

cannot be recognized easily which is disadvantage in

using thermal camera. The heat in the vehicle is

preserved in warm conditions while it diminishes

faster in cold conditions. Therefore, vehicles can be

recognized for a longer period of time in warmer

climate environments. It would be a challenge to

detect a vehicle without any heat in colder

environments based on the results. However, the

results might vary with larger training dataset. A

thermal camera does not work as a normal camera,

therefore, any algorithm or deep learning network

should be trained with the images or videos obtained

from the thermal camera which is also evident based

on the results obtained from the pre-trained detectors.

Training a deep learning network with many layers

take considerable amount of time. It took

approximately 8 hours to train the modified

FasterRCNN deep learning network with a single

CPU. The training time can be reduced with the use

of a graphical processing unit. There is also difference

between successful detection rates of vehicles and

empty spaces. In snowy conditions, vehicles which

are warm are successfully detected while few of the

empty spaces which are occluded are not detected

even in bright conditions. The lines of empty spaces

were not visible during snowy or dark conditions

making the detection of empty spaces challenging.

The empty spaces in third and fourth rows are

occluded and could not be detected in winter evening

or dark conditions.

6 CONCLUSION

The paper aims to identify parking occupancy

using a thermal camera. The first four rows of the

parking lot were considered to identify parking

occupancy. Pre-trained detectors, ACF detector,

HOG based cascade detector, Faster RCNN deep

network and modified Faster RCNN deep learning

network algorithms were used to identify parking

occupancy. The modified Faster RCNN deep learning

network performed better compared to other

detectors. Even with limited number of training

images, modified deep learning network achieved

88% successful detection of vehicles while it

achieved 66% of detection rate of empty spaces. The

future work can be focused on improving the

detection rate and to acquire real time parking

occupancy detection using thermal camera. This

paper addresses the first step in identifying parking

occupancy information and the next step would be to

train the detector with additional training images and

get the positional information of the vacant parking

space which can be fed to a parking guidance system.

REFERENCES

ACFDETECTOR. 2018. vehicleDetectorACF [Online].

Mathworks. Available: https://www.mathworks.com/

help/driving/ref/vehicledetectoracf.html [Accessed 25

Dec 2018].

Amato, G., Carrara, F., Falchi, F., Gennaro, C., Meghini, C.

& Vairo, C. 2017. Deep learning for decentralized

parking lot occupancy detection. Expert Systems with

Applications, 72, 327-334.

Bao, P., Zhang, L. & Wu, X. 2005. Canny edge detection

enhancement by scale multiplication. IEEE

transactions on pattern analysis and machine

intelligence, 27, 1485-1490.

Dalal, N. & Triggs, B. Histograms of oriented gradients for

human detection. Computer Vision and Pattern

Recognition, 2005. CVPR 2005. IEEE Computer

Society Conference on, 2005. IEEE, 886-893.

DATAINSPEKTIONEN. 2018. The purposes and scope of

the General Data Protection Regulation [Online].

Available: https://www.datainspektionen.se/other-

lang/in-english/the-general-data-protection-regulation-

gdpr/the-purposes-and-scope-of-the-general-data-

protection-regulation/ [Accessed].

De Almeida, P. R., Oliveira, L. S., Britto Jr, A. S., Silva Jr,

E. J. & Koerich, A. L. 2015. PKLot–A robust dataset

for parking lot classification. Expert Systems with

Applications, 42, 4937-4949.

Detector, C. 2018. Cascade detector [Online]. Available:

https://www.mathworks.com/help/vision/ug/train-a-

cascade-object-detector.html [Accessed].

Parking Occupancy Detection using Thermal Camera

489

Dollár, P., Appel, R., Belongie, S. & Perona, P. 2014. Fast

feature pyramids for object detection. IEEE

Transactions on Pattern Analysis and Machine

Intelligence, 36, 1532-1545.

Fan, Q., Brown, L. & Smith, J. A closer look at Faster R-

CNN for vehicle detection. Intelligent Vehicles

Symposium (IV), 2016 IEEE, 2016. IEEE, 124-129.

FASTERRCNN. 2018. vehicleDetectorFasterRCNN

[Online]. Mathworks. Available: https://

www.mathworks.com/help/driving/ref/vehicledetector

fasterrcnn.html [Accessed 20 Dec 2018].

Fleyeh, H. & Mohammed, I. 2012. Night Time Vehicle

Detection. Journal of Intelligent Systems, 21, 143-165.

Gaszczak, A., Breckon, T. P. & Han, J. Real-time people

and vehicle detection from UAV imagery. Intelligent

Robots and Computer Vision XXVIII: Algorithms and

Techniques, 2011. International Society for Optics and

Photonics, 78780B.

Hassoune, K., Dachry, W., Moutaouakkil, F. & Medromi,

H. Smart parking systems: A survey. Intelligent

Systems: Theories and Applications (SITA), 2016 11th

International Conference on. IEEE, 1-6.

Idris, M., Leng, Y., Tamil, E., Noor, N. & Razak, Z. 2009.

Саг park system: a review of smart parking system and

its technology. Information Technology Journal, 8,

101-113.

Iwasaki, Y., Misumi, M. & Nakamiya, T. 2013. Robust

vehicle detection under various environmental

conditions using an infrared thermal camera and its

application to road traffic flow monitoring. Sensors, 13,

7756-7773.

Liu, K. & Mattyus, G. 2015. Fast Multiclass Vehicle

Detection on Aerial Images. IEEE Geosci. Remote

Sensing Lett., 12, 1938-1942.

Maini, R. & Aggarwal, H. 2009. Study and comparison of

various image edge detection techniques. International

journal of image processing (IJIP), 3, 1-11.

Mao, L., Xie, M., Huang, Y. & Zhang, Y. Preceding vehicle

detection using Histograms of Oriented Gradients.

Communications, Circuits and Systems (ICCCAS),

2010 International Conference on. IEEE, 354-358.

Paidi, V., Fleyeh, H., Håkansson, J. & Nyberg, R. G. 2018.

Smart parking sensors, technologies and applications

for open parking lots: a review. IET Intelligent

Transport Systems, 12, 735-741.

Polycarpou, E., Lambrinos, L. & Protopapadakis, E. Smart

parking solutions for urban areas. World of Wireless,

Mobile and Multimedia Networks (WoWMoM), 2013

IEEE 14th International Symposium and Workshops on

a, 2013. IEEE, 1-6.

RESNET50. 2018. resnet50 [Online]. Available: https://

www.mathworks.com/help/deeplearning/ref/resnet50.

html?searchHighlight=resnet50&s_tid=doc_srchtitle

[Accessed].

Robert, K. Night-time traffic surveillance: A robust

framework for multi-vehicle detection, classification

and tracking. Advanced Video and Signal Based

Surveillance, 2009. AVSS'09. Sixth IEEE International

Conference on, 2009. IEEE, 1-6.

Shoup, D. C. 2006. Cruising for parking. Transport Policy,

13, 479-486.

Villegas, O. O. V., Balderrama, D. G., Domínguez, H. &

Sánchez, V. G. C. 2009. License plate recognition using

a novel fuzzy multilayer neural network. International

journal of computers, 3, 31-40.

Xu, Y., Yu, G., Wang, Y., Wu, X. & Ma, Y. 2016. A hybrid

vehicle detection method based on viola-jones and

HOG+ SVM from UAV images. Sensors, 16, 1325.

VEHITS 2019 - 5th International Conference on Vehicle Technology and Intelligent Transport Systems

490