Critical Vehicle Detection for Intelligent Transportation Systems

Erkut Akdag

∗

, Egor Bondarev and Peter H. N. De With

Department of Electrical Engineering, Eindhoven University of Technology, Eindhoven 5612AZ, The Netherlands

Keywords:

Vehicle Detection, Vehicle Classiﬁcation, Vision in Transformer, CNN, Intelligent Transportation.

Abstract:

An intelligent transportation system (ITS) is one of the core elements of smart cities, enhancing public safety

and relieving trafﬁc congestion. Detection and classiﬁcation of critical vehicles, such as police cars and

ambulances, passing through roadways form crucial use cases for ITS. This paper proposes a solution for

detecting and classifying safety-critical vehicles on urban roadways using deep learning models. At present,

a large-scale dataset for critical vehicles is not publicly available. The appearance scarcity of emergency

vehicles and different coloring standards in various countries are signiﬁcant challenges. To cope with the

mentioned drawbacks and to address the unique requirements of our smart city project, we ﬁrst generate a

large-scale critical vehicle dataset, combining images retrieved from various sources with the support of the

YOLO vehicle detection model. The classes of the generated dataset are: ﬁre truck, police car, ambulance,

military police car, dangerous truck, and standard vehicle. Second, we compare the performance of the Vision

in Transformer (ViT) network against the traditional convolutional neural networks (CNNs) for the task of

critical vehicle classiﬁcation. Experimental results on our dataset reveal that the ViT-based solution reaches

an average accuracy and recall of 99.39% and 99.34%, respectively.

1 INTRODUCTION

Surveillance cameras are increasingly adopted to ob-

serve trafﬁc ﬂow in public places to improve public

safety and reduce trafﬁc congestion. In the event of

high travel demand, or a dangerous trafﬁc incident,

trafﬁc congestion is likely to happen, especially in ur-

ban places. For instance, at large intersections or in

cities with a high trafﬁc density, accidents arise fre-

quently. In case of an accident, a rapid medical trans-

fer of the affected people may be necessary, where

even a short delay can be lethal. Considerable delay

in the ﬁrst response emergency services can occur due

to various reasons. For example, emergency vehicles,

such as ambulances, ﬁre trucks, and police cars, get

occasionally trapped in trafﬁc jams, thereby increas-

ing overall response time. In addition to the accidents,

trucks with dangerous cargo in city centers threaten

trafﬁc ﬂow and public safety. To address these prob-

lems, intelligent transportation systems (ITS) propose

beneﬁcial solutions via automated trafﬁc regulation

and notiﬁcation systems. Detection and tracking of

critical vehicles (emergency vehicles and trucks with

dangerous loads) become an essential part of such a

solution.

∗

Corresponding Author

Artiﬁcial neural networks (ANN), support vector

machines (SVM), and deep learning studies mainly

concentrate on regular vehicle classes, such as bus,

auto, cyclist, and van. At present, there is limited re-

search work on critical vehicle detection. The existing

studies focus on subsets of emergency vehicles, i.e.

excluding trucks carrying dangerous goods or mili-

tary police cars. However, inclusion of these classes

in the critical vehicle landscape is crucial. Critical ve-

hicles in this paper are a combination of emergency

vehicles (ambulances, ﬁre trucks, police cars, and

military police cars) and trucks with dangerous goods,

named “dangerous trucks“ throughout the paper.

This study aims at developing models for the crit-

ical vehicle detection problem, based on state-of-the-

art networks (e.g., YOLOv5, ViT, EfﬁcientNet, and

ResNet) as a ﬁrst step. To this end, we construct a

large-scale dataset by the YOLOv5 detection model

and label correction. As a second step, we gener-

ate critical vehicle models by applying three differ-

ent classiﬁcation architectures: EfﬁcientNet, ResNet-

50, and ViT, and evaluate the experimental results.

All models are publicly available (Wightman, 2019;

Glenn Jocher, 2021).

This paper is organized as follows. Section 2 pro-

vides an overview of the related work, including vehi-

Akdag, E., Bondarev, E. and N. De With, P.

Critical Vehicle Detection for Intelligent Transportation Systems.

DOI: 10.5220/0010968900003191

In Proceedings of the 8th International Conference on Vehicle Technology and Intelligent Transport Systems (VEHITS 2022), pages 165-171

ISBN: 978-989-758-573-9; ISSN: 2184-495X

165

cle detection and classiﬁcation for different use cases.

Section 3 explains the dataset creation and model gen-

eration approaches. The experimental results are eval-

uated in Section 4, while section 5 concludes the pa-

per.

2 RELATED WORK

This section presents a literature overview on the ve-

hicle detection and classiﬁcation problem. First, the

existing deep learning models are introduced for the

object detection task. Second, the studies targeting

the detection and classiﬁcation of regular vehicles are

discussed. Last, an overview is provided of existing

studies on the detection and classiﬁcation of emer-

gency vehicles.

2.1 Deep Learning Methods for Object

Detection

Deep neural network solutions are widely exploited

in visual recognition tasks such as detection and clas-

siﬁcation. The power of deep neural networks comes

from the hierarchical representation of the features,

which can be developed during the model training

phase. CNNs are able to learn and extract rich and

meaningful features that can be utilized for various

visual recognition tasks, such as classiﬁcation, image

segmentation, and object detection.

In the early studies on object detection, the

region-based convolutional neural network (RCNN)

model (Girshick et al., 2014) was proposed. The per-

formance of the RCNN object detector can be im-

proved signiﬁcantly by applying heavier and deeper

CNN models at the feature extraction stage, such

as the VGG network (Simonyan and Zisserman,

2014) and the residual network called ResNet-50 (He

et al., 2016), containing 50-layers. After the emer-

gence of the YOLO architecture, urban trafﬁc solu-

tions associated with object detection have switched

to YOLO models, since the inference time of the

YOLOv3 model is faster than the RCNN model. The

work in (Redmon and Farhadi, 2018) concludes that

YOLOv3 is one of the fastest models for object de-

tection, where it can process 45 images per sec-

ond. Therefore, in real-time applications, YOLO or

the single-shot detector (SSD) (Liu et al., 2016) ap-

proaches are predominantly considered. In the litera-

ture, YOLOv3 (Redmon and Farhadi, 2018) is exten-

sively practiced for object detection in surveillance,

including the vehicle detection problem.

2.2 Detection of Regular Vehicle Types

The detection and classiﬁcation of regular vehicles,

such as cars, buses, and trucks, are mainly studied by

performing experiments on different regular vehicle

datasets.

In the study by (Arinaldi et al., 2018), the au-

thors implement two different methods for the de-

tection and classiﬁcation of regular vehicles. The

ﬁrst method is a combination of Mixture of Gaussian

(MoG) (Reynolds, 2009) and support vector machine

(SVM) (Suykens and Vandewalle, 1999), while the

second one is FasterRCNN (Ren et al., 2015). Faster-

RCNN outperforms MoG in vehicle detection, and it

further surpasses SVM for the task of classifying the

vehicle types. Authors report results on the Indone-

sian Toll Road dataset (Arinaldi et al., 2018), and

the public MIT trafﬁc dataset (Wang et al., 2008),

designed to perform trafﬁc scene analysis and study

crowded scenes. In (Suhao et al., 2018), authors

concentrate on regular vehicle classes such as car,

minibus, and SUV, not covering any critical vehicles.

The FasterRCNN (Ren et al., 2015) architecture is ap-

plied to perform the vehicle detection and classiﬁca-

tion task. The VGG16 model is practiced to detect ve-

hicle types in various trafﬁc scenarios. The MIT (Dol-

lar et al., 2011) and Caltech car datasets (Kobayashi

et al., 2007) are utilized for model generation. In the

work by (Shekade et al., 2020), the authors deploy a

dataset with regular vehicles, which involves eleven

types of vehicles including bicycle, bus, car, truck,

and motorbike. In that research, the YOLOv3 ap-

proach is proposed for the vehicle detection task. In

the experiments, the total variety of vehicle motion at

a speciﬁc time is examined with the support of count-

ing and classifying vehicles. Authors conduct their

experiments on the MIO vision trafﬁc dataset (Jung

et al., 2017).

2.3 Detection of Emergency Vehicle

Types

The task of emergency vehicle detection and classiﬁ-

cation imposes two speciﬁc challenges. Emergency

vehicle datasets are scarce compared to the regular

vehicle datasets. It is challenging to gather data for

the emergency vehicle classes due to their rare occur-

rence on the public roads. Another challenge is that

the visual appearances of emergency vehicles vary be-

tween countries, for instance, the stripes or colors can

be different in each country.

One of the studies focusing on emergency vehi-

cles is performed by (Roy and Rahman, 2019). The

authors propose a system to detect emergency vehi-

VEHITS 2022 - 8th International Conference on Vehicle Technology and Intelligent Transport Systems

166

Resnet-50

ViT-base

EfficientNet

Dataset

Model Generation

Standard cars

Fire trucks

Ambulances

Dangerous trucks

Police cars

Military police cars

Stanford Cars Dataset

Object Detection

Flickr Dataset

Dataset Creation

YOLOv5

Model Evaluation

Figure 1: Proposed approach for critical vehicle detection, consisting of the vehicle sample generation from CCTV cameras,

YouTube videos, concatenation with Flickr and Stanford Cars dataset, followed by model generation and evaluation with

ResNet-50, ViT-base and EfﬁcientNet models.

cles in trafﬁc congestion, which helps the control sys-

tem to handle trafﬁc ﬂow and prioritize emergency

vehicles. In this study, vehicles are ﬁrst detected

from CCTV footage using a deep convolutional net-

work, and they are classiﬁed afterwards. YOLOv3

is selected for the vehicle detection task to catego-

rize truck, bus, and car classes, as provided in the

COCO dataset (Lin et al., 2014). These classes are

passed through the VGG-16 classiﬁer to identify the

vehicle type to be emergency or not. Authors con-

duct their experiments on the Stanford University cars

dataset (Szegedy et al., 2015) for regular vehicles

and the images collected from the internet for emer-

gency vehicles. Another study by (Carvalho Bar-

bosa et al., 2020) concentrates on the detection and

classiﬁcation of a few emergency vehicles, such as

ambulances, ﬁre trucks, and police cars, not includ-

ing military police cars and dangerous trucks. The

authors propose a model named priority vehicle im-

age detection network (PVIDNet) based on YOLOv3

by applying a soft-root-sign (SRS) activation func-

tion to decrease the execution time of the proposed

model. A database of Brazilian vehicle images is de-

ployed (Carvalho Barbosa et al., 2020) for testing.

As can be derived, previous studies mainly focus

on the detection and classiﬁcation of regular vehicles,

and only a few works focus on emergency vehicles. In

existing research on emergency vehicles, datasets are

limited, and multi-label classiﬁcation is rarely stud-

ied. We aim to generate a large-scale critical vehicle

dataset to advance further in this problem domain. A

detailed explanation of our dataset is given in Sec-

tion 3.

3 APPROACH

Figure 1 illustrates the proposed critical vehicle de-

tection approach in two steps, namely the dataset cre-

ation and model generation.

3.1 Dataset Creation

For model development and testing, the gener-

ated large-scale dataset includes several open-source

datasets and images retrieved from CCTV recordings.

First, we adopt the YOLOv5 model instead of

YOLOv3 used in previous studies to enhance reliable

and robust object detection. Signiﬁcant differences

between YOLOv5 and its prior releases are mosaic

data augmentation and auto-learning bounding-box

anchors. Likewise, YOLOv5 is nearly 90% lighter

than YOLOv4 (Bochkovskiy et al., 2020) in terms of

computational cost, while maintaining the detection

accuracy and still providing the competitive results.

Apart from that, YOLOv5 is one of the fastest deep

learning models according to EfﬁcientDet (Tan et al.,

2020). Experiments on a GTX 1080 GPU show that

YOLOv5 achieves an inference time average of 106

frames/second.

Our dataset includes emergency vehicles and dan-

gerous trucks having a rounded cylindrical shape and

transporting dangerous goods, such as chemicals and

gasoline. A subset of the Flickr dataset makes a

signiﬁcant contribution to all classes of the created

dataset. Moreover, we have searched among vari-

ous YouTube trafﬁc videos for obtaining Dutch emer-

gency vehicle classes, such as ambulance, police, and

Critical Vehicle Detection for Intelligent Transportation Systems

167

Table 1: Sample volumes for each class in the dataset generated from the Flickr, Stanford cars, YouTube, and CCTV sources.

Data sources Generated data

Class name Flickr Stanford Cars YouTube CCTV Total Train Test

Standard cars 1,343 5,000 - 102 6,445 5,156 1,289

Fire trucks 1,087 - - - 1,087 870 217

Ambulances 200 - 166 70 431 345 86

Police cars 621 - 85 30 736 589 147

Military police cars 363 - 117 - 480 384 96

Dangerous trucks 1,245 - - - 1,245 996 249

Standard cars Fire trucks Ambulances Dangerous trucks Police cars Military police cars

Figure 2: Samples of different vehicle classes in the generated dataset from left to right: standard cars, ﬁre trucks, ambulances,

dangerous trucks, police cars, and military police cars.

military police. We apply the YOLOv5 model on

the selected videos to increase the number of sam-

ples for the speciﬁed emergency classes in our gener-

ated dataset. To extend the scale of the standard car

class, we have combined the publicly available Stan-

ford University cars dataset with our dataset. Besides,

we have carefully examined open-source CCTV cam-

eras at different locations of the Netherlands to re-

trieve as many critical vehicle images as possible.

The above-mentioned CCTV video recordings pro-

vide further samples of standard cars, ambulances,

and police cars through applying the YOLOv5 model

for extraction and selection of car instances. The

dataset is completed by applying vehicle ﬁltering and

annotation processes on the generated image samples

for all classes.

Finally, the generated dataset includes 10,424 im-

ages classiﬁed into 6 categories: standard cars, ﬁre

trucks, ambulances, police cars, military police cars,

and dangerous trucks. The sample volumes of each

class are detailed and listed in Table 1. Additionally,

various samples of different vehicle classes are illus-

trated in Figure 2. We have split the generated dataset

into 80% for training and 20% for testing purposes.

Furthermore, standard data augmentation techniques

are applied to enlarge the available dataset and im-

prove our model.

3.2 Model Generation

Having the large-scale dataset of critical vehicles

available, we experiment with three architectures

(ViT, EfﬁcientNet, and ResNet) to design an ac-

curate multi-class critical vehicle detection model.

The transformer architecture has become the de-

facto standard for natural language processing ap-

plications. It outperforms other architectures, such

as LSTMs (Greff et al., 2016) and gated recurrent

units (GRUs) (Dey and Salem, 2017). As opposed to

CNNs, a transformer can be applied to the sequence

of image patches in the image classiﬁcation tasks. In

the ViT model, images are split into patches like to-

kens in an NLP application, and sequences of lin-

ear embeddings provide the input to the transformer

VEHITS 2022 - 8th International Conference on Vehicle Technology and Intelligent Transport Systems

168

Table 2: Comparison of image classiﬁcation performance metrics recall, accuracy, F

score for six vehicle classes on the

EfﬁcientNet, ViT-base, and ResNet-50 models and the metric averages.

EfﬁcientNet ViT-base ResNet-50

Class name Recall Accuracy F

Score Recall Accuracy F

Score

Standard cars 100.00 99.92 99.96 99.92 99.84 99.88 99.76 99.92 99.84

Fire trucks 100.00 99.54 99.77 100.00 100.00 100.00 100.00 99.54 99.77

Ambulances 98.83 100.00 99.41 100.00 100.00 100.00 98.83 96.59 97.70

Police cars 95.91 97.24 96.57 98.63 98.63 98.63 98.63 96.66 97.64

Military police cars 94.79 94.79 94.79 97.91 97.91 97.91 94.79 97.84 96.29

Dangerous trucks 99.59 99.20 99.39 99.59 100.00 99.79 99.59 100.00 99.79

Average 98.18 98.44 98.31 99.34 99.39 99.35 98.60 98.27 98.36

EfficientNet ViT-base ResNet-50

Figure 3: Confusion matrices combining the ground truth and predicted labels for each class using the EfﬁcientNet, ViT-base

and ResNet-50 models.

encoder. Positional information is also retained by

adding the position embeddings, which are learned

without hard-coded vector positions. Lastly, the se-

quences of the patch and positional embedding vec-

tors are supplied to the input stage of the transformer

encoder. There is also a special token at the start of

the ViT model, similar to bidirectional encoder rep-

resentations from transformers (BERT) (Devlin et al.,

2018). We practice the VITbase-21k model trained on

Imagenet-21k. Furthermore, we experiment with the

EfﬁcientNet and ResNet-50 models, which are two

state-of-the-art CNN models commonly used for the

classiﬁcation.

We train the models by hyperparameter tuning,

where throughout our experiments, we have used an

initial learning rate of 0.0002, reduced by one-tenth

at every 3 epochs. All models have been trained

for a total of 12 epochs using the stochastic gradi-

ent descent (SGD) algorithm and the PyTorch frame-

work (Paszke et al., 2019).

4 RESULTS AND EVALUATION

To evaluate the performance of the three models, we

deploy the following common metrics: recall, accu-

racy, and F

score. The F

score is an appropriate

metric for a class-imbalanced dataset similar to ours,

since it is the harmonic mean of the recall and preci-

sion. It presents the balance between how many ex-

amples of actual vehicles are classiﬁed correctly and

how likely the classiﬁcation is correct. A commonly

used strategy for performing experimental tests is to

combine pre-training with ﬁne-tuning to better lever-

age from the limited amount of vehicle images and

improve the classiﬁcation performance.

The results are summarized in Table 2. In this

table, we present six rows for each of the vehicle

classes. Three primary columns (EfﬁcientNet, ViT-

base, and ResNet-50) show the performance of each

corresponding classiﬁcation model, indicating recall,

accuracy, and F

score metrics. From the experimen-

tal results shown in Table 2, the average accuracy of

the ViT model can reach 99.39%, while EfﬁcientNet

and ResNet-50 reach 98.44% and 98.27%, respec-

Critical Vehicle Detection for Intelligent Transportation Systems

169

tively. Although the CNN models have better results

in a few class types, the ViT model achieves the best

experimental results for all metrics on average. Ex-

perimental results on the ambulance and ﬁre truck

classes are outstanding, since their color and stripes

are salient and distinguishable. Likewise, the met-

rics on the standard car class are high, since regular

cars do not have speciﬁc striping or coloring patterns.

However, these results are slightly lower for military

police and police cars. In terms of the F

score metric,

military police cars achieve the lowest total score with

94.79% in EfﬁcientNet and 96.29% in ResNet-50.

The class of police cars reveals the second-lowest ac-

curacy score with 97.24% in EfﬁcientNet and 96.66%

in ResNet-50. This is mainly because the shapes and

colors of both classes are almost identical from the

front view. The only difference in appearance is in

the back view of the car, where military police cars

have the grey or blue color and police cars have a

white color. The dominant performance of the ViT-

base model can be observed for all classes, especially

improving the classiﬁcation performance for the mil-

itary police and police cars.

Moreover, the confusion matrices for each classi-

ﬁcation model are presented in Figure 3. The vertical

axis shows the actual labels of corresponding vehicles

classes, while the horizontal axis shows the correctly

predicted number of vehicles for each class. The con-

fusion matrices shown in Figure 3 indicate the source

of the problem with police and military police cars.

The ViT model accomplishes this difﬁcult task in con-

trast to other models.

Furthermore, we have compared the latency

for each classiﬁcation model. The latency is a

processing-time measurement to determine the per-

formance of various models for a speciﬁc applica-

tion, where the latency is deﬁned as the time re-

quired for processing one image. The ViT model

takes 0.017 seconds to classify one image, while Efﬁ-

cientNet and ResNet-50 consume 0.015 seconds and

0.013 seconds, respectively. While the ViT model

performs better than other models for the critical ve-

hicle classiﬁcation task, it reaches a slightly higher

latency, which is acceptable.

5 CONCLUSION

This work studies the critical vehicle detection prob-

lem in trafﬁc for safety reasons, which is rarely ad-

dressed in the literature. First, we have carefully gath-

ered all vehicle classes from trafﬁc surveillance data

and various datasets, to generate a large-scale crit-

ical vehicle database. The created dataset includes

10,424 images classiﬁed into 6 categories: standard

cars, ﬁre trucks, ambulances, police cars, military po-

lice cars, and dangerous trucks. Second, we have

employed the ViT and the EfﬁcientNet and ResNet-

50 CNN models to experiment and identify the most

suitable model for the task. We have chosen the ViT

model because it achieves excellent accuracy for each

critical vehicle class. Detailed experimental results

show that the ViT-based solution reaches an average

accuracy higher than 0.99. Furthermore, experimental

latency measurements indicate that all three models

have similar values. All the trained models used for

the proposed approach are publicly available (Wight-

man, 2019), (Glenn Jocher, 2021).

ACKNOWLEDGEMENTS

This work is supported by the European ITEA project

SMART on intelligent trafﬁc ﬂow systems and Efﬁ-

cient Deep Learning (EDL) RMR project.

REFERENCES

Arinaldi, A., Pradana, J. A., and Gurusinga, A. A. (2018).

Detection and classiﬁcation of vehicles for trafﬁc

video analytics. Procedia computer science, 144:259–

268.

Bochkovskiy, A., Wang, C.-Y., and Liao, H.-Y. M. (2020).

Yolov4: Optimal speed and accuracy of object detec-

tion. arXiv preprint arXiv:2004.10934.

Carvalho Barbosa, R., Shoaib Ayub, M., Lopes Rosa, R.,

Zegarra Rodr

ıguez, D., and Wuttisittikulkij, L. (2020).

Lightweight pvidnet: A priority vehicles detection

network model based on deep learning for intelligent

trafﬁc lights. Sensors, 20(21):6218.

Devlin, J., Chang, M.-W., Lee, K., and Toutanova, K.

(2018). Bert: Pre-training of deep bidirectional trans-

formers for language understanding. arXiv preprint

arXiv:1810.04805.

Dey, R. and Salem, F. M. (2017). Gate-variants of gated

recurrent unit (gru) neural networks. In 2017 IEEE

60th international midwest symposium on circuits and

systems (MWSCAS), pages 1597–1600. IEEE.

Dollar, P., Wojek, C., Schiele, B., and Perona, P. (2011).

Pedestrian detection: An evaluation of the state of the

art. IEEE transactions on pattern analysis and ma-

chine intelligence, 34(4):743–761.

Girshick, R., Donahue, J., Darrell, T., and Malik, J. (2014).

Rich feature hierarchies for accurate object detec-

tion and semantic segmentation. In Proceedings of

the IEEE conference on computer vision and pattern

recognition, pages 580–587.

Glenn Jocher, e. a. (2021). ultralytics/yolov5: v6.0

- YOLOv5n ’Nano’ models. https://github.com/

ultralytics/yolov5.

VEHITS 2022 - 8th International Conference on Vehicle Technology and Intelligent Transport Systems

170

Greff, K., Srivastava, R. K., Koutn

ık, J., Steunebrink, B. R.,

and Schmidhuber, J. (2016). Lstm: A search space

odyssey. IEEE transactions on neural networks and

learning systems, 28(10):2222–2232.

He, K., Zhang, X., Ren, S., and Sun, J. (2016). Deep resid-

ual learning for image recognition. In Proceedings of

the IEEE conference on computer vision and pattern

recognition, pages 770–778.

Jung, H., Choi, M.-K., Jung, J., Lee, J.-H., Kwon, S., and

Young Jung, W. (2017). Resnet-based vehicle classiﬁ-

cation and localization in trafﬁc surveillance systems.

In Proceedings of the IEEE conference on computer

vision and pattern recognition workshops, pages 61–

67.

Kobayashi, T., Hidaka, A., and Kurita, T. (2007). Selection

of histograms of oriented gradients features for pedes-

trian detection. In International conference on neural

information processing, pages 598–607. Springer.

Lin, T.-Y., Maire, M., Belongie, S., Hays, J., Perona, P.,

Ramanan, D., Doll

ar, P., and Zitnick, C. L. (2014).

Microsoft coco: Common objects in context. In Euro-

pean conference on computer vision, pages 740–755.

Springer.

Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, S.,

Fu, C.-Y., and Berg, A. C. (2016). Ssd: Single shot

multibox detector. In European conference on com-

puter vision, pages 21–37. Springer.

Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J.,

Chanan, G., Killeen, T., Lin, Z., Gimelshein, N.,

Antiga, L., et al. (2019). Pytorch: An imperative style,

high-performance deep learning library. Advances in

neural information processing systems, 32.

Redmon, J. and Farhadi, A. (2018). Yolov3: An incremental

improvement. arXiv preprint arXiv:1804.02767.

Ren, S., He, K., Girshick, R., and Sun, J. (2015). Faster

r-cnn: Towards real-time object detection with region

proposal networks. Advances in neural information

processing systems, 28:91–99.

Reynolds, D. A. (2009). Gaussian mixture models. Ency-

clopedia of biometrics, 741:659–663.

Roy, S. and Rahman, M. S. (2019). Emergency vehicle de-

tection on heavy trafﬁc road from cctv footage using

deep convolutional neural network. In 2019 Interna-

tional Conference on Electrical, Computer and Com-

munication Engineering (ECCE), pages 1–6. IEEE.

Shekade, A., Mahale, R., Shetage, R., Singh, A., and

Gadakh, P. (2020). Vehicle classiﬁcation in trafﬁc

surveillance system using yolov3 model. In 2020 In-

ternational Conference on Electronics and Sustain-

able Communication Systems (ICESC), pages 1015–

1019. IEEE.

Simonyan, K. and Zisserman, A. (2014). Very deep con-

volutional networks for large-scale image recognition.

arXiv preprint arXiv:1409.1556.

Suhao, L., Jinzhao, L., Guoquan, L., Tong, B., Huiqian,

W., and Yu, P. (2018). Vehicle type detection based

on deep learning in trafﬁc scene. Procedia computer

science, 131:564–572.

Suykens, J. A. and Vandewalle, J. (1999). Least squares

support vector machine classiﬁers. Neural processing

letters, 9(3):293–300.

Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S.,

Anguelov, D., Erhan, D., Vanhoucke, V., and Rabi-

novich, A. (2015). Going deeper with convolutions.

In Proceedings of the IEEE conference on computer

vision and pattern recognition, pages 1–9.

Tan, M., Pang, R., and Le, Q. V. (2020). Efﬁcientdet: Scal-

able and efﬁcient object detection. In Proceedings

of the IEEE/CVF conference on computer vision and

pattern recognition, pages 10781–10790.

Wang, X., Ma, X., and Grimson, W. E. L. (2008). Un-

supervised activity perception in crowded and com-

plicated scenes using hierarchical bayesian models.

IEEE Transactions on pattern analysis and machine

intelligence, 31(3):539–555.

Wightman, R. (2019). Pytorch image models. https:

//github.com/rwightman/pytorch-image-models.

Critical Vehicle Detection for Intelligent Transportation Systems

171