Unconstrained License Plate Detection in Hardware

Petr Musil

, Roman Jur

anek

and Pavel Zem

ık

FIT, Brno University of Technology, Brno, Czech Republic

Keywords:

ALPR, Soft Cascade, Decision Trees, WaldBoost.

Abstract:

In this paper, we propose an FPGA implementation of license plate detection (LPD) in images captured by

arbitrarily placed cameras, vehicle-mounted cameras, or even handheld cameras. In such images, the license

plates can appear in a wide variety of positions and angles. Thus we cannot rely on a-priori known geometric

properties of the license plates as many contemporary applications do. Unlike the existing solutions targeted

for DSP, FPGA or similar low power devices, we do not make any assumptions about license plate size and

orientation in the image. We use multiple sliding window detectors based on simple image features, each

tuned to a speciﬁc range of projections. On a dataset captured by a camera mounted on a vehicle, we show

that detection rate is 98 % (and 98.7 % when combined with video tracking). We demonstrate that our FPGA

implementation can process 1280×1024 pixel image at over 40 FPS with a minimum width of detected license

plates approximately 100 pixels. The FPGA block is fully functional and it is intended to be used in a smart

camera to parking control in residential zones.

1 INTRODUCTION

License plate detection (LPD) is an essential part of

applications, such as detection of vehicles for trafﬁc

monitoring and enforcement purposes, or as a basis

for automatic license plate recognition. In many in-

dustrial use cases, LPD is solved at a sufﬁciently ad-

vanced level in scenarios with restricted size, orienta-

tion, and in environment with controlled illumination

(e.g. using IR ﬂashes). A vast number of research pa-

pers has addressed this controlled scenario, and it is

far beyond the scope of this paper to thoroughly re-

view them.

In this paper, we focus on license plate detection

implemented in FPGA in an image captured by arbi-

trarily placed cameras, vehicle-mounted cameras, or

even handheld cameras. The target applications in-

clude parking control in residential zones, detection

of stolen vehicles, and other applications, in which

license plates of cars are automatically detected and

recognized. Our goal was to improve the performance

of license plate detection process, to implement the

method in FPGA, and to provide the solution that

requires low resources and that is suitable for inte-

gration into hardware devices, or even directly into

https://orcid.org/0000-0001-9766-4759

https://orcid.org/0000-0003-0589-0172

https://orcid.org/0000-0001-7969-5877

37.27

42.33

35.77

32.09

Figure 1: We detect license plates in wide variety of defor-

mations in challenging light conditions. The images show

the detected license plates with a conﬁdence value.

cameras. The license plates captured in the image

are distorted from their original rectangular shape by

perspective projection, see Figure 1 for a few exam-

ples. Due to the character of the data, the detection

and localization can not rely on an a-priori informa-

tion, such as license plate size or orientation in the

image. Moreover, lighting conditions, ranging from

Musil, P., Juránek, R. and Zem

cík, P.

Unconstrained License Plate Detection in Hardware.

DOI: 10.5220/0010174000130021

In Proceedings of the 7th International Conference on Vehicle Technology and Intelligent Transport Systems (VEHITS 2021), pages 13-21

ISBN: 978-989-758-513-5

Table 1: Summary of properties of state of the art detection algorithms. The accuracies reported by the works are not directly

comparable as the methods were evaluated on different data and by different testing protocols.

Rotation Candidate search Detection Platform Accuracy

(Mai et al., 2011) tilt Vertical gradient detection Character recognition PC 98 %

(Yang et al., 2011) Chroma and gradient ﬁltering SVM FPGA/PC N/A

(Zhai et al., 2011) Edge detection, CCA Filtering by size, orientation and aspect ratio FPGA 99 %

(Jeffrey and Ramalingam, 2012) Edge and background detection, CCA Gradient statistics PC 97 %

(Khalil and Kurniawan, 2014) Edge detection SVM PC 88 %

(Biyabani et al., 2015) Edge detection Segmentation with projections, template matching FPGA 84 %

(Ha and Shakeri, 2016) Edge and background detection Character recognition with template matching PC 84 %

(Chhabra et al., 2016) Segmentation with edge detection Filtering by size and aspect ratio FPGA 99 %

(Elbamby et al., 2016) Edge detection LBP cascade PC 94 %

(Yuan et al., 2017) Adaptive threshold and edge detection SVM PC 96 %

(Xie et al., 2018) Free – CNN–YOLO PC/GPU 99.5 %

(Silva and Jung, 2018) Free – CNN PC/GPU 98.35 %

(Sborz et al., 2019) Morphologic operation, CCA Character segmentation FPGA –

(Youseﬁ et al., 2019) Background subtraction LBP cascade PC 98 %

(Chen and Wang, 2020) Free – CNN + AdaBoos cascade PC/GPU 99.9 %

(Gao et al., 2020) Free – CNN PC/GPU 99.9 %

(WU et al., 2020) corner detection and morphological op. CNN PC/GPU 97.2 %

This work Free – soft cascade with LBP FPGA 98 %

deep shadows to overexposures, prevents using edge

detection, gradient-based methods, and segmentation-

based methods.

2 RELATED WORK

An overview of several works on license plate detec-

tion with a focus on those implemented in embed-

ded systems or PC is shown in Table 1. Many of the

works are intended for stationary cameras and appli-

cations where the rough location and size of license

plate is known (Biyabani et al., 2015; Ha and Shak-

eri, 2016; Jeffrey and Ramalingam, 2012; Zhai et al.,

2011; Chhabra et al., 2016; Khalil and Kurniawan,

2014; Yang et al., 2011; Yuan et al., 2017; Elbamby

et al., 2016; Hsu et al., 2013). In many cases, they

rely on the presence of the license plate edges to re-

duce search space.

Many methods (Anagnostopoulos et al., 2006;

Zhai et al., 2011; Jeffrey and Ramalingam, 2012)

are based on connected component analysis (CCA)

for the search of the candidate components in bina-

rized images. (Zhai et al., 2011) uses CCA with mor-

phological operators and ﬁlters the components using

knowledge about aspect ratio, size, and orientation.

(Anagnostopoulos et al., 2006) segments characters

within the CCA components by horizontal and verti-

cal projections. (Mai et al., 2011) uses edge-based

approach and rotation-free character recognition to

search for tilted license plates. (Wang and Lee, 2003)

uses CCA in combination with Radon transform to

search for components with license plate properties

(size, aspect ratio). (Sborz et al., 2019) proposed li-

cense plate localization on FPGA. They use two mor-

phological operations and connected components la-

belling technique to ﬁnd license plate candidates.

Template matching-based methods (Ha and Shak-

eri, 2016; Biyabani et al., 2015) use background sub-

traction and edge detection for candidate search. Can-

didates are then scanned for characters using template

matching. Other methods (Chhabra et al., 2016; Biya-

bani et al., 2015) use template matching to ﬁnd license

plate candidates, and searches for characters in hori-

zontal and vertical projections of the license plate im-

ages.

Machine learning based methods for licence plate

detection typically use either SVM (Khalil and Kur-

niawan, 2014; Yang et al., 2011; Yuan et al., 2017),

cascade classiﬁers (Elbamby et al., 2016; Arth et al.,

2006) or Convolutional neural network(CNN)(Xie

et al., 2018; Chen and Wang, 2020; Gao et al., 2020;

WU et al., 2020). (Khalil and Kurniawan, 2014)

searches for candidate locations by horizontal and

vertical edge detection and classiﬁes them by SVM

with co-occurrence matrix features. (Yang et al.,

2011) presented a hybrid FPGA-PC approach where

an image is ﬁltered by chroma ﬁlter and gradient ﬁlter

on FPGA, and then SVM is applied on candidate po-

sitions on PC. (Yuan et al., 2017) also proposed adap-

tive thresholding, density ﬁlter and SVM with colour

saliency features.

(Arth et al., 2006) proposes a Haar cascade detec-

tor on DSP for detection of cars and license plates.

(Elbamby et al., 2016) uses cascade classiﬁer with

Local Binary Patterns (LBP) features and prepro-

cesses the image by edge detection and morphol-

ogy ﬁlters. (Youseﬁ et al., 2019) use AdaBoost cas-

cade with LBP features for license plate detection and

background subtraction for non-moving areas elimi-

nation. They use linear regression for estimation of

size of detected plates. (Xie et al., 2018) use YOLO

(You only look once) method (Redmon et al., 2015)

based on CNN to multi-directional license plate de-

tection. (Chen and Wang, 2020) proposed a method

combing the CNN with broad learning system based

VEHITS 2021 - 7th International Conference on Vehicle Technology and Intelligent Transport Systems

scaling

Figure 2: (left) Image scales and objects detected by the classiﬁers. (right) Final detections after non-maximum suppression.

The classiﬁers are color-coded, the yellow one detects license plates almost aligned, the green one detects slightly tilted

license plates.

on AdaBoost for license plate recognition. (Gao et al.,

2020) created CNN based encoder-decoder system

where encoder detects candidate license plate char-

acters and recognises them without considering the

format of the license plate. (WU et al., 2020) uses

colour segmentation, corner detection and morpho-

logical operations to select the operating area, and ﬁ-

nally uses a convolutional neural network to get li-

cense plate area.

Generic object detector hardware architectures

have also been published, mostly demonstrated on

face detection (Cho et al., 2009; Kyrkou and

Theocharides, 2011; Zemcik et al., 2013; Said and

Atri, 2016; Musil et al., 2020), that could be applied

for detection of license plates. In most cases, they

implement Haar Cascades (Viola and Jones, 2004) or

SVM (Dalal and Triggs, 2005).

In this work, similarly to (Arth et al., 2006; El-

bamby et al., 2016), we propose statistical sliding

window detectors trained by the machine learning al-

gorithm. Such an approach is more robust than pure

image processing, such as in (Anagnostopoulos et al.,

2006; Zhai et al., 2011), since it easily adapts to visual

variability of data. The difference in our approach

from the other ones is that we use multiple detectors

tuned to various deformations of license plates in or-

der to boost the quality of detection. We use a large

artiﬁcially generated dataset of training examples and

propose an FPGA implementation of the detection

process on Xilinx Zynq platform. Our architecture is

based on (Musil et al., 2020) which provides excellent

speed/resource tradeoff. We evaluated the method on

a large image and video database obtained from in-

dustry.

To our knowledge, this work is the ﬁrst to use

boosted classiﬁers for unconstrained detection of li-

cense plates directly in FPGA. Other works that ex-

poit FPGA use mostly segmentation or ﬁltering ap-

proaches which are targeted for speciﬁc scenarios

(and usually fail in unconstraned detection). Recent

works employ mostly neural networks which are su-

perior in accuracy of detection but their implementa-

tion in FPGA is limited and expensive (Nguyen et al.,

2019; Wu et al., 2019) or they require speciﬁc hard-

ware (GPU, VPU) (Xie et al., 2018; Chen and Wang,

2020; Gao et al., 2020; WU et al., 2020).

3 LICENSE PLATE

LOCALIZATION

Our task is to detect license plates observed by a cam-

era and to localize them; see Figure 1 for example

images. The intended applications require high ac-

curacy of the detection result. At the same time, it

can tolerate a reasonable amount of false detections

(which can be ﬁltered out later, for example, by the

license plate recognition process). Our license plate

detection algorithm is based on multiple independent

boosted classiﬁers, whose results are merged to get

the ﬁnal results. Each of the classiﬁers captures li-

cense plates with a speciﬁc range of observed in-plane

rotations. We use constant soft cascade classiﬁers

(

Sochman and Matas, 2005; Doll

ar et al., 2014) with

Local Binary Patterns (LBP) features (Zhang et al.,

2007; Zemcik et al., 2013). The detection process is

illustrated in Figure 2,

3.1 The Classiﬁer

The classiﬁer is a function H(x) which gives the con-

ﬁdence value for an image patch x, formally deﬁned

in Equation (1). During the detection process, every

location of the input image is analyzed by the classi-

ﬁer. Multi-scale detection is solved by image scaling

by a ﬁxed factor. The classiﬁer H(x), is a sequence of

T weak classiﬁcation functions (T = 512 in our ex-

periments), and its response on image window x is a

sum of predictions produced by the individual weak

classiﬁers. We use simple weak classiﬁers based on

Unconstrained License Plate Detection in Hardware

Table 2: Ranges of orientation and window sizes for three classiﬁers.

Range [

◦

] Window [px] LP images

0 < φ < 15 18×64

10 < φ < 30 22×56

25 < φ < 45 28×43

LBP features f and lookup tables A with the conﬁ-

dence predictions.

(x) =

∑

i=1

( f

(x)) (1)

The detection process on an image I produces a

set of locations and sizes that were not rejected by the

classiﬁer, Equation (2). Each candidate comprises of

its location x, y in the image, size w, h and conﬁdence

score c (the classiﬁer response on the image patch cor-

responding to the location).

H(I) =

{

(x, y, w, h, c)

}

(2)

An important property of soft cascade classiﬁers

is that the classiﬁcation function H(x) can be ter-

minated after evaluating k-th weak classiﬁer, when

(x) < θ

. Thresholds θ are trained, so that ma-

jority of background samples is rejected early in the

process. The computational complexity of the clas-

siﬁer dramatically decreases compared to the case of

evaluation of all T weak classiﬁers, and it can be eval-

uated as an average number of weak classiﬁers eval-

uated per image position nf (which is usually orders

of magnitude lower that T ). This value is especially

important as it directly inﬂuences the speed of the de-

tector in the FPGA implementation.

3.2 Detection with Multiple Classiﬁers

We propose to use n classiﬁers H =

(1)

, . . . , H

(n)

for the localization of license plates, each tuned for

LPs with speciﬁc range of rotation angle φ. Each clas-

siﬁer produces detections independently, and the ﬁnal

set of candidates is obtained as a union of all candi-

dates (3). Final locations are produced by a simple,

overlap-based non-maxima suppression algorithm.

H (I) =

[

(i)

(I) (3)

Each detector is trained for a speciﬁc range of li-

cense plate angles φ. The range assigned to the k-th

classiﬁer is,

max



45(k − 1)

− 5, 0)



< φ <

45k

where 45

◦

marks the upper limit of license plate ori-

entations. The classiﬁer window aspect ratio is set

as a mean aspect ratio of license plates falling in the

range φ. The window size is set to constant area of

1024 pixels and 2 pixel margin is added. The ranges

and window sizes for n = 3 classiﬁers are summarized

in Table 2.

The number of classiﬁers in the ensemble H inﬂu-

ences the accuracy and speed of the detection process.

One classiﬁer can not capture all the variability of the

license plates, while more detectors are more accu-

rate but slower. This trend is shown in Figure 3. Two

classiﬁers give already reasonable detection rate and

sufﬁcient speed. We observed that using more classi-

ﬁers results in better localization. Using more classi-

ﬁers, however, reduces speed, and for this reason, we

use three classiﬁers in our hardware implementation

(Section 4).

The training data we used is a broad set of license

plate images randomly transformed to match defor-

mations likely to occur in the target application, since

the scenario is often known and ﬁxed, e.g. in case of

our vehicle-mounted camera. Few samples are shown

in Table 2. The advantage of such an approach is that

it can be automated and any number of training sam-

ples with precise ground truth (including orientation)

can be quite effortlessly generated for each applica-

tion case.

VEHITS 2021 - 7th International Conference on Vehicle Technology and Intelligent Transport Systems

# classifiers

1 2 3 4 5

Miss rate

0.05

0.1

(a)

# classifiers

1 2 3 4 5

(b)

# classifiers

1 2 3 4 5

Speed [FPS]

100

(c)

Figure 3: Effect of the number of the classiﬁers on the detection. ((a)) Reduced miss rate, ((b)) higher computational com-

plexity and ((c)) lower frame rate. We choose to use 3 classiﬁers which is a good balance between detection accuracy, speed

and hardware requirements (see Section 4).

4 FPGA ARCHITECTURE

We implemented the LP detector on our hardware

camera platform based on SoC Xilinx Zynq XC

7Z020. The SoC contains two ARM CortexA9 cores

and FPGA interconnected through high-speed AMBA

AXI bus. The approach is, however, quite general

and can be applied to almost any Xilinx Zynq family

member and also to other platforms. In our case, the

platform is equipped with a low noise global shutter

CMOS image sensor Python1300 from ON Semicon-

ductor, which is attached directly to FPGA, forming

a smart camera. The sensor resolution is 1280×1024

pixels and it can capture up to 210 FPS.

Figure 4 shows the block diagram of the camera

with the detector integrated into it. As the minimum

size of the license plate needed to detect is approxi-

mately 100 pixels in width, we downscale the input

image to half resolution before the detection phase.

The image is passed to the detector block through

balancing FIFO that covers irregularities in the detec-

tor speed that depend on the image content and that

are hard to predict. The results of detection – loca-

tions of detected objects – are transmitted to ARM

CPU memory, along with the original image. On the

CPU, the detected license plates are tracked using the

Kalman ﬁlter. The camera outputs cropped license

plate images that are transmitted to the server for fur-

ther processing (OCR, storage, etc.).

The Detector block is based on (Musil et al.,

2020) which implements multi-scale soft cascade de-

tector with LBP (Zhang et al., 2007) or LRD features

(Hradi

s et al., 2008). The engine works as a pro-

grammable automaton, using a sequence of feature

parameters as instructions. We modiﬁed the engine in

order to use three classiﬁers and adapted it to a more

recent platform with more resources and memory.

The engine works without external memory, stor-

ing only a narrow stripe of the image which is be-

ing stored and analyzed directly in BRAM inside the

FPGA. The stripe memory size is 4096×32 pixels in

32 BRAMs organized in a way that data for a fea-

Table 3: FPGA resource utilization.

BRAM LUT REG

Image acquisition 5 2559 4746

Balancing FIFO 4 115 210

Detector 38 9521 7099

Storage 5 3734 4785

Total 52 15938 16840

7Z020 res. 37 % 30 % 16 %

ture evaluation can be obtained in one clock cycle.

The stripe memory is ﬁlled from the CMOS, and the

scaled versions of the image are created on-the-ﬂy;

the process is illustrated in Figure 4.

The classiﬁcation function is executed overall po-

sitions in the stripe memory. The processing is heav-

ily pipelined, we use two pipelines of length 9, so up

to 18 positions are processed in parallel. The efﬁ-

ciency of feature extraction is np = 1.75 features ex-

tracted in clock cycle. Overall performance in frames

per second, expressed by equation (4), is, in general,

dependent on the clock frequency f and the total num-

ber of features that must be calculated on an image

for all classiﬁers (i.e. the number of positions P times

the average number of features n f for each individual

classiﬁer).

F =

f · np

∑

i=1

· n f

(4)

In this work we assume f = 200MHz; P

872487, P

= 899067, and P

= 915442; and n f

2.89, n f

= 3.23, and n f

= 2.78. The values of n f

were estimated on a testing set of 1522 images and

are valid for the experimental classiﬁers presented in

the Section 5. The estimated upper limit of perfor-

mance of the detection unit is therefore F = 48 FPS.

Table 3 summarizes resources required by the design.

The whole solution takes approximately one third of

resources of the 7Z020 FPGA which we use in our

solution.

Unconstrained License Plate Detection in Hardware

Image

correction

Scale

FIFO

Detector

DMA

VDMA

ARM

Memory

Data from

sensor

Stripe memory 4096x32 px

Scale + FIFO

Current line 640 px

Data from

sensor

6x6

5x5

Scaling

...

Figure 4: (Top) Block scheme of the camera prototype. (Bottom) The data from sensor are passed to a narrow image stripe

in memory. All scales are created automatically.

5 DETECTOR EVALUATION

Figure 5: Recall-Precision characteristics for the proposed

method (for different n) and comparison with other meth-

ods.

We evaluated the proposed method in two modes –

Single image mode and Tracking mode.

In the Single image mode we evaluate precision-

recall trade-off on a large number of testing images.

However, it is impossible to compare to other pub-

lished license plate detection methods (summarized in

Table 1) since they published results on different test-

ing datasets using different testing protocols. More-

over, none of them has available source code. For this

reasons, we compare our method to well known gen-

eral object detection methods – ACF (Doll

ar et al.,

2014) and MTCNN (Zhang et al., 2016), and a com-

mercial solution Plate Recognizer

. It must be how-

ever noted that none of them is targeted for FPGA

with all of its constraints (low memory, integer arith-

metic, etc.). MTCNN is a recent method based on

www.platerecognizer.com provides free REST API

Table 4: Tracking evaluation on testing sequences.

Seq. #tracked #missed #false Det. rate

1 197 1 59 0.995

2 103 1 43 0.990

3 101 5 15 0.953

4 181 1 26 0.995

5 205 3 23 0.986

6 53 0 2 1.000

Total 840 11 168 0.987

neural networks and it require PC/GPU platform.

Plate Recognizer internal structure was not published.

ACF is a method similar to ours since it shares a com-

mon classiﬁer structure and the detection algorithm.

For the evaluation, we used our internal dataset of

1 522 images with 811 manually annotated license

plate bounding boxes. The results are summarized in

Figure 5. Not surprisingly, MTCNN and the com-

mercial solution give almost perfect results. ACF is

comparable to our method in terms of recall but with

higher precision. The results for our method show

that a single classiﬁer already gives reasonable results

and adding more classiﬁers tuned to different trans-

formations further improves them (see also Figure 3).

Our method can reach recall over 95% with preci-

sion around 0.8, and recall 98% with precision 0.7

which is sufﬁcient for real-world applications. In

the Tracking mode, we detect and track LPs in video

sequences and evaluate the detection rate. We used a

simple tracker based on Kalman ﬁlter. In the evalua-

tion, we required each license plate to be hit at least

once in sequence. From the application point of view,

the Tracking mode is more important. In the test,

we use six sequences, each approximately 10 min-

utes long taken at ten frames per second. The tracking

accuracy is summarized in Table 4. The total detec-

VEHITS 2021 - 7th International Conference on Vehicle Technology and Intelligent Transport Systems

26.61

44.32

28.85

52.43

39.6

47.89

31.95

47.55

15.94

38.39

61.11

39.28

Figure 6: (top) Examples of detected license plates and detector responses. (bottom) Missed license plates (annotations

marked by red).

tion rate 98.7 %. Missed license plates are in most

cases, those with extreme deformations or very dirty,

false alarms are license plate-like patterns in the im-

age (various signs, etc.), see Figure 6 for a few exam-

ples.

6 CONCLUSIONS

We presented a method for reliable real-time detec-

tion and localization of license plates observed from

a moving vehicle, and its FPGA implementation suit-

able for integration into a smart camera. Our solu-

tion is unique, since it uses machine learning-based

detection method, it does not require external mem-

ory, and enables for real-time frame rates on low-end

FPGA. In this paper, we focused on the detection of

license plates. However, the method is more general

and can be adapted for detection of other objects too.

The proposed solution is capable of handling complex

projections of the license plates, such as deformations

caused by perspective projection, scaling, and rota-

tions. The proposed solution uses multiple, in our

case three, classiﬁers to ensure the quality of the re-

sults while keeping the performance high. The classi-

ﬁers are based on simple LBP features which makes it

easy to implement them in FPGA. We demonstrated

that our hardware solution could process 1280×1024

pixel frames at over 40 FPS while taking only a frac-

tion of resources of low-end FPGA (Zynq Z7020).

The accuracy of detection measured on the real-world

data is 98 % and 98.7 % when combined with track-

ing, while producing only a modest amount of false

detections.

Future work includes further efﬁciency improve-

ments, investigation of dependency of quality on free-

dom of projection of images and sampling of the im-

age frames, and possibly also exploitation of other

features and detection mechanisms including hybrid

approach with neural networks.

ACKNOWLEDGEMENTS

This work is part of the FitOptiVis project funded

by the ECSEL Joint Undertaking under grant number

H2020-ECSEL-2017-2-783162.

REFERENCES

Anagnostopoulos, C. N. E., Anagnostopoulos, I. E.,

Loumos, V., and Kayafas, E. (2006). A license

plate-recognition algorithm for intelligent transporta-

tion system applications. IEEE Transactions on Intel-

ligent Transportation Systems, 7(3):377–392.

Arth, C., Bischof, H., and Leistner, C. (2006). TRICam –

an embedded platform for remote trafﬁc surveillance.

In 2006 Conference on Computer Vision and Pattern

Recognition Workshop (CVPRW’06), pages 125–125.

Biyabani, A. A., Al-Salman, S. A., and Alkhalaf, K. S.

(2015). Embedded real-time bilingual alpr. In Com-

munications, Signal Processing, and their Applica-

tions (ICCSPA).

Unconstrained License Plate Detection in Hardware

Chen, C. L. P. and Wang, B. (2020). Random-positioned

license plate recognition using hybrid broad learning

system and convolutional networks. IEEE Transac-

tions on Intelligent Transportation Systems, pages 1–

13.

Chhabra, S., Jain, H., and Saini, S. (2016). Fpga based

hardware implementation of automatic vehicle license

plate detection system. In 2016 International Con-

ference on Advances in Computing, Communications

and Informatics (ICACCI), pages 1181–1187.

Cho, J., Mirzaei, S., Oberg, J., and Kastner, R. (2009).

Fpga-based face detection system using haar classi-

ﬁers. In FPGA.

Dalal, N. and Triggs, B. (2005). Histograms of oriented

gradients for human detection. In CVPR.

Doll

ar, P., Appel, R., Belongie, S., and Perona, P. (2014).

Fast feature pyramids for object detection. IEEE

Transactions on Pattern Analysis and Machine Intel-

ligence, 36(8):1532–1545.

Elbamby, A., Hemayed, E. E., Helal, D., and Rehan, M.

(2016). Real-time automatic multi-style license plate

detection in videos. In Computer Engineering Con-

ference (ICENCO).

Gao, F., Cai, Y., Ge, Y., and Lu, S. (2020). Edf-lpr: a new

encoder decoder framework for license plate recogni-

tion. IET Intelligent Transport Systems, 14(8):959–

969.

Ha, P. S. and Shakeri, M. (2016). License plate automatic

recognition based on edge detection. In 2016 Artiﬁcial

Intelligence and Robotics (IRANOPEN), pages 170–

174.

Hradi

s, M., Herout, A., and Zemcik, P. (2008). Local

rank patterns - novel features for rapid object detec-

tion. In Proceedings of International Conference on

Computer Vision and Graphics 2008, Lecture Notes

in Computer Science, pages 1–2.

Hsu, G., Chen, J., and Chung, Y. (2013). Application-

oriented license plate recognition. IEEE Transactions

on Vehicular Technology, 62(2):552–561.

Jeffrey, Z. and Ramalingam, S. (2012). High deﬁnition li-

cence plate detection algorithm. In Southeastcon.

Khalil, M. S. and Kurniawan, F. (2014). License plate detec-

tion method for real-time video of low-cost webcam

based on hybrid svm-heuristic approach. In Informa-

tion Technology: New Generations (ITNG).

Kyrkou, C. and Theocharides, T. (2011). A ﬂexible paral-

lel hardware architecture for adaboost-based real-time

object detection. In VLSI Systems.

Mai, V., Miao, D., Wang, R., and Zhang, H. (2011). An im-

proved method for vietnam license plate location. In

2011 International Conference on Multimedia Tech-

nology, pages 2942–2946.

Musil, P., Jur

anek, R., Musil, M., and Zem

ık, P. (2020).

Cascaded stripe memory engines for multi-scale ob-

ject detection in fpga. IEEE Transactions on Circuits

and Systems for Video Technology, 30(1):267–280.

Nguyen, D. T., Nguyen, T. N., Kim, H., and Lee, H. (2019).

A high-throughput and power-efﬁcient fpga imple-

mentation of yolo cnn for object detection. IEEE

Transactions on Very Large Scale Integration (VLSI)

Systems, 27(8):1861–1873.

Redmon, J., Divvala, S. K., Girshick, R. B., and Farhadi, A.

(2015). You only look once: Uniﬁed, real-time object

detection. CoRR, abs/1506.02640.

Said, Y. and Atri, M. (2016). Efﬁcient and high-

performance pedestrian detector implementation for

intelligent vehicles. IET Intelligent Transport Sys-

tems, 10(6):438–444.

Sborz, G. A. M., Pohl, G. A., Viel, F., and Zeferino, C. A.

(2019). A custom processor for an fpga-based plat-

form for automatic license plate recognition. In 2019

32nd Symposium on Integrated Circuits and Systems

Design (SBCCI), pages 1–6.

Silva, S. M. and Jung, C. R. (2018). License plate detection

and recognition in unconstrained scenarios. In 2018

European Conference on Computer Vision (ECCV),

pages 580–596.

Viola, P. and Jones, M. J. (2004). Robust real-time face

detection. IJCV.

Sochman, J. and Matas, J. (2005). WaldBoost – learning for

time constrained sequential detection. In CVPR.

Wang, S.-Z. and Lee, H.-J. (2003). Detection and recog-

nition of license plate characters with different ap-

pearances. In Proceedings of the 2003 IEEE Interna-

tional Conference on Intelligent Transportation Sys-

tems, volume 2, pages 979–984 vol.2.

Wu, D., Zhang, Y., Jia, X., Tian, L., Li, T., Sui, L., Xie, D.,

and Shan, Y. (2019). A high-performance cnn proces-

sor based on fpga for mobilenets. In 2019 29th In-

ternational Conference on Field Programmable Logic

and Applications (FPL), pages 136–143.

WU, X., QIU, J., and QIU, A. (2020). An efﬁcient li-

cense plate location algorithm based on deep leaming.

In 2020 International Conference on Computer Engi-

neering and Application (ICCEA), pages 543–546.

Xie, L., Ahmad, T., Jin, L., Liu, Y., and Zhang, S. (2018).

A new cnn-based method for multi-directional car li-

cense plate detection. IEEE Transactions on Intelli-

gent Transportation Systems, 19(2):507–517.

Yang, S. Y., Lu, Y. C., Chen, L. Y., and Cherng, D. C.

(2011). Hardware-accelerated vehicle license plate

detection at high-deﬁnition image. In 2011 First In-

ternational Conference on Robot, Vision and Signal

Processing, pages 106–109.

Youseﬁ, E., Nazem Deligani, A. H., Jafari Amirbandi, J.,

and Karimzadeh Kiskani, M. (2019). Real-time scale-

invariant license plate detection using cascade classi-

ﬁers. In 2019 IEEE Conference on Multimedia Infor-

mation Processing and Retrieval (MIPR), pages 399–

402.

Yuan, Y., Zou, W., Zhao, Y., Wang, X., Hu, X., and Ko-

modakis, N. (2017). A robust and efﬁcient approach

to license plate detection. IEEE Transactions on Im-

age Processing, 26:1102 – 1114.

Zemcik, P., Juranek, R., Musil, P., Musil, M., and Hradis,

M. (2013). High performance architecture for object

detection in streamed videos. In Field Programmable

Logic and Applications (FPL).

VEHITS 2021 - 7th International Conference on Vehicle Technology and Intelligent Transport Systems

Zhai, X., Bensaali, F., and Ramalingam, S. (2011). Real-

time license plate localisation on fpga. In CVPR 2011

WORKSHOPS, pages 14–19.

Zhang, K., Zhang, Z., Li, Z., and Qiao, Y. (2016). Joint

face detection and alignment using multitask cascaded

convolutional networks. IEEE Signal Processing Let-

ters, 23(10):1499–1503.

Zhang, L., Chu, R., Xiang, S., Liao, S., and Li, S. Z. (2007).

Face detection based on multi-block lbp representa-

tion. In ICB.

Unconstrained License Plate Detection in Hardware