Crowd Behavior Analysis based on Convolutional Neural Network:

Social Distancing Control COVID-19

Fatma Bouhlel

1 a

, Hazar Mliki

2 b

and Mohamed Hammami

3 c

MIRACL-FSEG, University of Sfax, Faculty of Economics and Management of Sfax, Road Airport Km 4, 3018 Sfax,

Tunisia

MIRACL-ENET’COM, University of Sfax, National School of Electronics and Telecommunications of Sfax,

Road Tunis City El Ons, 3018 Sfax, Tunisia

MIRACL-FS, University of Sfax, Faculty of Sciences of Sfax, Road Sokra Km 3, 3018 Sfax, Tunisia

Keywords:

COVID-19, Crowd Behavior, Social Distancing, Crowd Density Estimation, Human Detection, Convolutional

Neural Network, UAV.

Abstract:

The outbreak of the COVID-19 and the lack of pharmaceutical intervention increase the spread of COVID-19.

Since no vaccine or treatment are yet available, social distancing represents a good strategy to control the

propagation of this pandemic and learn to live with it. In this context, we introduce a new approach for crowd

behavior analysis from UAV-captured video sequences in order to monitor social distancing. The proposed

approach involves two methods: a macroscopic method and a microscopic method. The macroscopic method

aims to estimate the crowd density by classifying the aerial frame patches into four categories: Dense, Sparse,

Medium and None. However, the microscopic method allows to detect and track humans and then compute

the distance between them. The quantitative and qualitative results validate the performance of our methods

compared to the state-of-the-art references.

1 INTRODUCTION

In December 2019, Wuhan city which is the capital

of Hubei became the propagation source of a pneu-

monia outbreak, known as coronavirus disease 2019

(COVID-19) (Lewnard and Lo, 2020; Zhang et al.,

2020; Zhou et al., 2020). According to the num-

ber of conﬁrmed cases and the number of caused

death, the COVID-19 has declared by the World

Health Organization (WHO) as though a pandemic

virus (WHO, 2020). Indeed, this pandemic was prop-

agated promptly (Punn et al., 2020) to more than 216

countries. In fact, about 26.121.999 conﬁrmed cases

are noted along with 864.618 deaths in the world on

Septembre 4, 2020. The lack of active therapeutic

agents and the dearth of immunity raise the spread

of COVID-19 (Punn et al., 2020; Singh and Adhikari,

2020; Wilder-Smith and Freedman, 2020). As no vac-

cines are currently available, social distancing con-

stitutes an effective strategy allowing the standstill

of this pandemic spread (Punn et al., 2020; Singh

and Adhikari, 2020; Park et al., 2020; Cristani et al.,

2020). In fact, social distancing in pandemic time has

https://orcid.org/0000-0001-9979-9729

https://orcid.org/0000-0002-0285-0944

https://orcid.org/0000-0003-3580-0473

internal as well as external beneﬁts helping humans to

learn living with this virus (Zhou et al., 2020). The in-

ternal beneﬁts consist of being less probably to catch

the virus in the case of socially distant. As for exter-

nal beneﬁt, it helps reducing the propagation of the

virus to other people, precisely other strangers. In the

context of COVID-19, it is recommended to maintain

a distance of two meters from other humans (Repici

et al., 2020).

For this purpose, we introduce a new approach,

that tends to help the social inspection force by alert-

ing them in the case of non-compliance with social

distancing measures. The proposed approach aims to

analyze the crowd behavior from UAV-captured video

sequences in order to monitor social distancing. Our

approach consists of two methods: a macroscopic

method and a microscopic method. These methods

are based on the use of convolutional neural net-

works (CNN) and transfer learning. The macroscopic

method estimates the crowd density by classifying the

aerial frames patches within four categories: Dense,

Sparse, Medium and None. The microscopic method

detects and tracks humans, to compute the distance

between them. The main contributions of this paper

are outlined as follows:

• The proposed strategy of dividing the input frame

Bouhlel, F., Mliki, H. and Hammami, M.

Crowd Behavior Analysis based on Convolutional Neural Network: Social Distancing Control COVID-19.

DOI: 10.5220/0010193002730280

In Proceedings of the 16th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2021) - Volume 5: VISAPP, pages

273-280

ISBN: 978-989-758-488-6

273

into N-sub-frames and estimating their density in

parallel process allows to reduce the time com-

putation. Such contribution makes our approach

suitable for real-time applications.

• The complementary combination of macroscopic

and microscopic methods: The ﬁltering level pro-

cess in the macroscopic method allows selecting

the suspicous crowd regions and focuses on the

most critical situation. In fact, the detection of

dense or medium crowd regions will trigger an

alert to the inspection forces (police). However,

the detection of a sparse crowd regions, will ac-

tivate the microscopic process to check the main-

tain of the social distancing within these regions.

This ﬁltering process provides a particular focus

on the most suspicious crowd regions and reduces

the complexity cost and hence the time computa-

tion.

• Our approach stands out from the literature (Punn

et al., 2020) in the use of UAV sensor which di-

verts the ﬁxed coverage areas problem and as-

sure open spaces surveillance (Hazar et al., 2019;

Mliki et al., 2020). In fact, the UAV provides

more ﬂexibility enabling it to overcome the oc-

clusion problem and monitor open space areas.

The remaining sections of the paper are organized

as follows. Section 2 identiﬁed the related works on

macroscopic and microscopic crowd behavior analy-

sis. In section 3, the outlines of the proposed approach

are reported. The experimental results are discussed

in section 4. The last section reviews the proposed ap-

proach and offered some future research perspectives.

2 RELATED WORKS

The crowd behavior analysis is based on two types

of complementary analysis methods: macroscopic

method and microscopic method. In the following

sections, the related works of these two methods in

aerial views were explored.

2.1 Macroscopic Methods

The macroscopic methods handle visually indistin-

guishable crowded (VIC) aerial frames and analyze

the crowd as a global entity. In fact, these methods

neglect the local information and focus on the treat-

ment of global information such as crowd density,

crowd counting and crowd ﬂow. In this state-of-the-

art study, we are interested in crowd density estima-

tion methods in aerial views. In this context, Meyn-

berg and Kuschk (Meynberg and Kuschk, 2013) pro-

posed to encode crowd patches through the Gabor ﬁl-

ter bank. Next, the resulted Gabor features are fed to

the support vector machine (SVM) classiﬁer. Finally,

they classify crowd patches into two categories: dense

and none. For the purpose of improving their previous

work, the selfsame authors (Meynberg et al., 2016)

proposed to make use of an LBP descriptor instead

of the Gabor descriptor. They managed to estimate

the crowd density in a ﬁxed ground sampling distance

(GSD) aerial images which ranges from 10 to 13 cm.

Within the same axis, Mliki et al. (Hazar et al., 2019),

extracted the points of interest from each crowd patch

trough the scale-invariant feature transform descrip-

tor (SIFT). Afterward, they extracted from each in-

terest centered region the texture features using the

multiblock local binary pattern (MB-LBP) descriptor.

Thereafter, they generated a global MB-LBP feature

vector through the concatenated of each centered re-

gion MB-LBP vector features. Finally, the obtained

global MB-LBP features are fed to the SVM classi-

ﬁer.

2.2 Microscopic Methods

The microscopic methods analyze crowd behavior by

detecting each person in the crowd and analyzing its

behavior. These methods describe efﬁciently the un-

usual events in an aerial view. Nonetheless, they can

only handle the sparse crowds, as they are not able to

separate humans in dense or medium crowds. There-

fore, we analyzed person detection methods in aerial

videos. AlDahoul et al. (AlDahoul et al., 2018) pro-

posed a real-time human detection method from the

UAV-captured video sequences. As a pretreatment,

they used an optical ﬂow technique to detect poten-

tial motion regions. Next, they classiﬁed these poten-

tial motion regions using a pretrained CNN AlexNet

(Krizhevsky et al., 2012). Nonetheless, this method

doesn’t deal with the close objects constraint. To han-

dle this constraint, Mliki et al. (Mliki et al., 2020)

proposed to integrate a module of regions of inter-

est generation and selection which helps adapting the

classical CNN, devoted to the classiﬁcation problem,

to detect humans.

2.3 Discussion

In the literature, two types of complementary crowd

behavior analysis methods are identiﬁed: macro-

scopic method and microscopic method. Although

the good performance achieved by handcrafted meth-

ods (Hazar et al., 2019; Meynberg and Kuschk, 2013;

Meynberg et al., 2016), they overlook the semantic

information resulting from extremely abstract deep

VISAPP 2021 - 16th International Conference on Computer Vision Theory and Applications

274

features. Furthermore, they depend on the choice

of descriptor (Mliki et al., 2020). Referring to this

state-of-the-art study, we propose a new macroscopic

method for crowd density estimation based on pre-

trained CNN. Regarding the microscopic methods,

we addopted the Mliki et al. (Mliki et al., 2020)

method since it deals with the close objects constraint

and adapts the classical CNN, devoted to the classi-

ﬁcation problem, to detect humans. Nonetheless, we

proposed to add the human tracking step, in order to

generate continous humans location and then comput-

ing distance between them.

3 PROPOSED APPROACH

In order to monitor COVID-19 social distancing, we

proposed a new approach for crowd behavior analysis

from UAV-captured video sequences. The proposed

approach consists of two methods as shown in Fig. 1:

a macroscopic method and a microscopic method.

The macroscopic method aims to estimate the crowd

density by classifying the aerial frame patches into

four categories: Dense, Sparse, Medium and None.

The microscopic method allows to detect and track

humans in order to compute the distance between

them. In the following sections, each of these meth-

ods was detailed.

3.1 Macroscopic Method

Since the proposed approach is designed for real time

applications, it’s important to improve the computa-

tion time and the algorithm complexity. Therefore,

we parallelize the crowd density estimation process

on multiple CPU to reduce time computation. In

fact, each aerial frame is divided into N equal sub-

frames depending on the number of the CPUs Cores.

The crowd density estimation proposed method is

based on a pretrained CNN. Various pretrained CNN

have been proposed such as: AlexNet (Krizhevsky

et al., 2012), ZFNet (Zeiler and Fergus, 2014), VG-

GNet (Simonyan and Zisserman, 2014), GoogLeNet

(Szegedy et al., 2015), ResNet (He et al., 2016). Bas-

ing on a comparative study performed by Kaiming et

al. (He et al., 2016) which assesses the pre-trained

CNN effectiveness in terms of the number of lay-

ers and classiﬁcation error rate, we adopted the pre-

trained CNN ‘AlexNet’. AlexNet

Thereby, we substitute the classiﬁcation layer by a

novel softmax layer to classify the crowd patches into

four categories: Dense, Sparse, Medium and None.

Thereafter, we ﬁne-tuned the obtained model to

our context of study in order to generate an adequate

Figure 1: Proposed approach for crowd behavior analysis to

monitor social distancing.

Crowd Behavior Analysis based on Convolutional Neural Network: Social Distancing Control COVID-19

275

crowd density model able to ﬁlter crowd regions and

focuses on the most critical ones. Hence, three sce-

narios can be addressed as illustrated in Fig. 2:

Figure 2: Filtering level process.

• Crowd patch is classiﬁed as ‘none’: this patch is

empty.

• Crowd patch is classiﬁed as ‘dense’ or ‘medium’:

this patch presents non-conformity with social

distancing problem. Therefore, an alert is trig-

gered to the social inspection forces with GPS lo-

cation information.

• Crowd patch is classiﬁed as ‘sparse’: this patch is

suspicious and needs more focus to check the con-

formity of social distancing using a microscopic

method.

3.2 Microscopic Method

This method aims to focus on the ‘sparse’ crowd re-

gion where people are distinguishable. The proposed

method consists of 3 steps: (1) Moving objects de-

tection, (2) Human detection and (3) Social distance

computation.

3.2.1 Moving Objects Detection

This step allows eliminating the acquisition sensor

ego-motion and detecting the potential motion re-

gions. In fact, the estimation of each pixel motion

provides a set of motion vectors resulting from both

of the potential motion regions and the UAV displace-

ment in the scene. To distinguish these motions, the

speed of the motion included from UAV is assumed

to be lower than the speed of the potential motion re-

gions (Mliki et al., 2020). Therefore, the motion re-

lated to the UAV displacement, which affects dynamic

as well as static objects in the scene, is not taken into

consideration. As for the motion related to the ob-

jects, it is detected by computing the optical ﬂow us-

ing the Lucas-Kande algorithm (Thota et al., 2013)

thanks to its speed, simplicity and accuracy (Mliki

et al., 2020).

3.2.2 Human Detection

The human detection required the generation of

human/non-human model to classify potential motion

regions. We generated the human/non-human model

using the pretrained CNN AlexNet (AlDahoul et al.,

2018). In order to adapt the classic CNN to handle

the detection problem, we integrated a module of re-

gions of interest generation and selection. The regions

of interest generation is performed using the Edge

boxes (Zitnick and Doll

ar, 2014). Since, the gener-

ated regions vary slightly in shape, scale or position,

we performed on these regions a selection step using

the Non-Maximum Suppression(NMS) (Zitnick and

Doll

ar, 2014) algorithm. Such module allows han-

dling the close-up objects, which often appear in the

selfsame potential motion region and are usually clas-

siﬁed as a single object. Afterward, the selected re-

gions are classiﬁed into human/non-human objects.

Next the detected humans are tracked using Kalman

ﬁlter (Welch and Bishop, 1995), in order to get con-

tinuous human location.

3.2.3 Social Distance Computation

We computed the distance between the detected hu-

mans on each frame. According to the computed dis-

tance, we used a green bounding box on the detected

human if he is away at least two meters from all hu-

mans. Otherwise, we used a red bounding box and

an alert is triggered. The compute of social distance

is based on the estimation of the ground sampling dis-

tance (GSD), which deﬁnes an image pixel size on the

ground. Fig. 3 describes the Ground Sampling Dis-

tance parameters.

The GSD is computed through the following equa-

tions:

GSD

[cm/px] =

Altitude[cm] × Sensorheight[cm]

Focallength[cm] × Fr ameheight[px]

(1)

GSD

[cm/px] =

Altitude[cm] × Sensorwidth[cm]

Focallength[cm] × Fr amewidth[px]

(2)

In the case when UAV have usually square pixels,

the GSD = GSD

= GSD

. However, when GSD

GSD

, then GSD takes the maximum value. Then, we

extracted the centroid of each human bounding box

and we computed the Euclidean distance between the

VISAPP 2021 - 16th International Conference on Computer Vision Theory and Applications

276

Figure 3: Ground sampling distance parameters.

obtained centroids, as follows:

D(p

, p

)[px] =

(cx

− cx

)

+ (cy

− cy

)

(3)

Where, (cx

,cy

) and (cx

,cy

) are respectively

the coordinates of the centroid of the detected person

and p

. In order to convert the distance in centime-

ters, we multiply it by the GSD.

4 EXPERIMENTAL STUDY

In this section, we evaluated the performance of the

proposed approach for crowd behavior analysis from

UAV-captured video sequences in order to monitor

social distancing. In the following sections, we de-

scribed the used datasets and the experimental results.

4.1 Datasets Description

The assessment of the proposed macroscopic method

was achieved on Mayenberg’s dataset (Meynberg and

Kuschk, 2013; Meynberg et al., 2016) and Mliki’s

dataset (Hazar et al., 2019), while the evaluation of

the microscopic method was carried out on the UCF-

ARG dataset (Nagendran et al., 2010). These datasets

were captured by UAV in an controlled environment.

4.1.1 Mayenberg’s Dataset

The Mayenberg’s dataset (Meynberg and Kuschk,

2013; Meynberg et al., 2016), was captured, through

two ﬂight campaigns, in open-air rock Germany fes-

tivals. This dataset includes 70.000 patches cor-

responding to one of the crowd classes: Dense,

Medium, Sparse and None. It shows minor variations

in the UAV-captured crowd images.

4.1.2 Mliki’s Dataset

This dataset was collected over the net and it repre-

sents various contexts as: festival, marathon, mani-

festation and pilgrimage (Hazar et al., 2019). It is

composed of 12,000 UAV-captured patches belong to

the crowd classes: Dense, Medium, Sparse and None.

The Mliki’s dataset shows more complex variations

than Mayenberg’s dataset.

4.1.3 UCF-ARG Dataset

The UCF-ARG dataset (Nagendran et al., 2010), in-

cludes 482 video sequences of human activities and

it deals with several constraints such as: the variation

of the UAV altitude, the severe ego-motion, the varia-

tion in the point of view and the illumination variation

conditions.

4.2 Experimental Results

We carried out two series of experiments: The ﬁrst se-

ries of experiments aimed to assess the macroscopic

method for crowd density estimation, and the sec-

ond series of experiments evaluated the microscopic

method for human detection and social distance com-

puting.

4.2.1 First Series of Experiments: Crowd

Density Estimation (Macroscopic Method)

Through this series of experiments, we assessed the

performance of the proposed macroscopic method for

crowd density classiﬁcation with the state-of-the-art

reference methods (Hazar et al., 2019; Meynberg and

Kuschk, 2013; Meynberg et al., 2016). For the pur-

pose of assuring fair comparison with Mayenberg’s

(Meynberg and Kuschk, 2013; Meynberg et al., 2016)

and Mliki’s (Hazar et al., 2019) methods, we applied

the same experimental protocol. In fact, we used, for

each dataset, 110 training crowd patches and 500 test-

ing crowd patches per class. Table 1 reported this

evaluation study in terms of accuracy rate.

Referring to this study, we noted that our results

surpassed those obtained by (Meynberg and Kuschk,

2013; Meynberg et al., 2016) owing to the use of

CNN and transfer learning, which deal with the UAV

constraints like the variation in: the GSD, the resolu-

tion, the illumination and the view angle. Nonethe-

less, Mliki et al. (Hazar et al., 2019) slightly exceed

our method, this is justiﬁed by the fact of applying

a three-level classiﬁcation strategy to deal with the

Crowd Behavior Analysis based on Convolutional Neural Network: Social Distancing Control COVID-19

277

Table 1: Comparison of our macroscopic method for crowd

density estimation with the state-of-the-art reference meth-

ods in terms of accuracy rate.

Datasets Methods Accuracy

Mayenberg

Meynberg et al., 2013 62.3 %

Mayenberg et al., 2016 74.67 %

Mliki et al., 2019 89.2 %

Our method 84.85 %

Mliki

Mayenberg et al., 2016 71.8%

Mliki et al., 2019 86.8 %

Our method 82.1 %

crowd density class confusion. Although this strategy

enhanced the performence results, it has signiﬁcantly

increased the computational complexity making it not

suitable for real-time contexet.

Table 2 illustrates the run-time gain of our macro-

scopic method while using N parallel process. Such

gain is critical for real-time applications.

Table 2: Evaluation of the parallel process for crowd den-

sity estimation on Mliki’s dataset (Hazar et al., 2019) in

terms of sec/frame.

Material conﬁguration Execution Sec/frame

i5 2.4 GHZ CPU (N=2)

8 GB RAM

Sequential 6.40

Parallel 5.60

Figure 4 represents some samples results of our

macroscopic method on the Mliki’s dataset (Hazar

et al., 2019). More qualitative results of our macro-

scopic method are available on this link: video1.

4.2.2 Second Series of Experiments: Human

Detection and Distance Computation

(Microscopic Method)

Through this series of experiments, we evaluated

the performance of the proposed microscopic method

for human detection. For fair comparison with Al-

Dahoul et al. (AlDahoul et al., 2018), we used their

experimental protocol. Table 3 highlighted this com-

parative study performed on the UCF-ARG dataset

(Nagendran et al., 2010).

Table 3: Comparison of our microscopic method for human

detection with (AlDahoul et al., 2018) method in terms of

accuracy rate.

Methods Accuracy

Al-Dahoul et al., 2018 98.0907%

Our method 99.5873 %

Through this experimental study, we underline

that our method outperformed the obtained results of

the AlDahoul et al. (AlDahoul et al., 2018) in terms of

accuracy rate. Such results highlighted the contribu-

tion of using the regions of interest generation and se-

lection module, which overcame the problem of close

objects classiﬁcation. Figure 5 shows some results

of our social distance computing on the UCF-ARG

dataset (Nagendran et al., 2010), where the red boxes

indicate the non-conformity of people to the social

distancing rule and green boxes refer to the respect

of the social distancing rule.

Further qualitative results of our microscopic

method are available on this link: video2.

5 CONCLUSION

The social distancing represents is the most recom-

manded strategy that allows the mitigating of the

COVID-19 pandemic propagation. In this context, we

have proposed a new approach that aims to analyze

crowd behavior from UAV-captured video sequences

to monitor social distancing. Our approach involves

two methods: a macroscopic method and a micro-

scopic method. The macroscopic method allows es-

timating the crowd density by classifying the aerial

frames patches into four categories: Dense, Sparse,

Medium and None. The microscopic method aims to

detect and track humans so as to compute the distance

between them. Through the quantitative and qualita-

tive evaluations of the proposed methods, we prove

the performance of our approach compared to the ex-

isting ones.

As future perspectives, we aim to improve the pro-

posed crowd behavior analysis approach by adding

a person re-identiﬁcation step in order to re-identify

persons who are non-compliance with social distanc-

ing measures. Furthermore, we are planning to inte-

grate the person temperature measurement by using

thermal UAV.

VISAPP 2021 - 16th International Conference on Computer Vision Theory and Applications

278

Figure 4: Crowd density estimation samples using the proposed macroscopic method on the Mliki’s dataset (Hazar et al.,

2019).

Figure 5: Humans detection, tracking and social distance monitoring samples using the proposed macroscopic method on the

UCF-ARG dataset (Nagendran et al., 2010).

Crowd Behavior Analysis based on Convolutional Neural Network: Social Distancing Control COVID-19

279

REFERENCES

AlDahoul, N., Md Sabri, A. Q., and Mansoor, A. M. (2018).

Real-time human detection for aerial captured video

sequences via deep models. Computational intelli-

gence and neuroscience, 2018.

Cristani, M., Del Bue, A., Murino, V., Setti, F., and Vincia-

relli, A. (2020). The visual social distancing problem.

arXiv preprint arXiv:2005.04813.

Hazar, M., Arous, O., and Hammami, M. (2019). Abnormal

crowd density estimation in aerial images. Journal of

Electronic Imaging, 28(1):013047.

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.

Krizhevsky, A., Sutskever, I., and Hinton, G. E. (2012). Im-

agenet classiﬁcation with deep convolutional neural

networks. In Advances in neural information process-

ing systems, pages 1097–1105.

Lewnard, J. A. and Lo, N. C. (2020). Scientiﬁc and ethical

basis for social-distancing interventions against covid-

19. The Lancet. Infectious diseases, 20(6):631.

Meynberg, O., Cui, S., and Reinartz, P. (2016). Detection

of high-density crowds in aerial images using texture

classiﬁcation. Remote Sensing, 8(6):470.

Meynberg, O. and Kuschk, G. (2013). Airborne crowd den-

sity estimation. ISPRS Annals of the Photogramme-

try, Remote Sensing and Spatial Information Sciences,

2:49–54.

Mliki, H., Bouhlel, F., and Hammami, M. (2020). Hu-

man activity recognition from uav-captured video se-

quences. Pattern Recognition, 100:107140.

Nagendran, A., Harper, D., and Shah, M. (2010). UCF-

ARG dataset, university of central ﬂorida. http://crcv.

ucf.edu/data/UCF-ARG.php.

Park, S. W., Sun, K., Viboud, C., Grenfell, B. T., and

Dushoff, J. (2020). Potential roles of social distancing

in mitigating the spread of coronavirus disease 2019

(covid-19) in south korea. medRxiv.

Punn, N. S., Sonbhadra, S. K., and Agarwal, S. (2020).

Monitoring covid-19 social distancing with person de-

tection and tracking via ﬁne-tuned yolo v3 and deep-

sort techniques. arXiv preprint arXiv:2005.01385.

Repici, A., Maselli, R., Colombo, M., Gabbiadini, R.,

Spadaccini, M., Anderloni, A., Carrara, S., Fugazza,

A., Di Leo, M., and Galtieri, P. A. (2020). Coron-

avirus (covid-19) outbreak: what the department of

endoscopy should know. Gastrointestinal endoscopy.

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

volutional networks for large-scale image recognition.

arXiv preprint arXiv:1409.1556.

Singh, R. and Adhikari, R. (2020). Age-structured impact

of social distancing on the covid-19 epidemic in india.

arXiv preprint arXiv:2003.12055.

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.

Thota, S. D., Vemulapalli, K. S., Chintalapati, K., and

Gudipudi, P. S. S. (2013). Comparison between the

optical ﬂow computational techniques. International

Journal of Engineering Trends and Technology, 4(10).

Welch, G. and Bishop, G. (1995). An introduction to the

kalman ﬁlter.

WHO (Septembre 4, 2020). World health organization

corona-viruses (covid-19). https://covid19.who.int/.

Wilder-Smith, A. and Freedman, D. O. (2020). Isolation,

quarantine, social distancing and community contain-

ment: pivotal role for old-style public health measures

in the novel coronavirus (2019-ncov) outbreak. Jour-

nal of travel medicine, 27(2):taaa020.

Zeiler, M. D. and Fergus, R. (2014). Visualizing and under-

standing convolutional networks. In European confer-

ence on computer vision, pages 818–833. Springer.

Zhang, J., Litvinova, M., Liang, Y., Wang, Y., Wang, W.,

Zhao, S., Wu, Q., Merler, S., Viboud, C., and Vespig-

nani, A. (2020). Age proﬁle of susceptibility, mixing,

and social distancing shape the dynamics of the novel

coronavirus disease 2019 outbreak in china. medRxiv.

Zhou, F., Yu, T., Du, R., Fan, G., Liu, Y., Liu, Z., Xiang,

J., Wang, Y., Song, B., and Gu, X. (2020). Clinical

course and risk factors for mortality of adult inpatients

with covid-19 in wuhan, china: a retrospective cohort

study. The lancet.

Zitnick, C. L. and Doll

ar, P. (2014). Edge boxes: Locating

object proposals from edges. In European conference

on computer vision, pages 391–405. Springer.

VISAPP 2021 - 16th International Conference on Computer Vision Theory and Applications

280