Coronavirus spread limitation using detective smart system

Given the circumstances, the world is going through due to the novel coronavirus (COVID-19); this paper proposes a new smart system that aims to reduce the spread of the virus. The proposed COVID-19 containment system is designed to be installed outside hospitals and medical centers. Additionally, it works at night as well as at daylight. The system is based on deep learning applied to pedestrian temperature data sets that are collected using thermal cameras. The data set is primarily of the temperature of pedestrians around medical centers. The thermal cameras are paired with conventional cameras for image capturing and cross-referencing the target pedestrian with an existing central database (Big Data). If the target is positive, the system sends a text message to the potentially infected person's cell phone upon recognition. The advisory sent text may contain useful information such as the nearest testing or isolation facility. This proposed system is assumed to be linked with the bigger network of the country’s COVID-19 response efforts. The simulation results reveal that the system can achieve an average precision of 90% fever detection among pedestrians.


Introduction
A preprint has previously been published (Morsy et al. 2021).Coronavirus is a family of viruses (SARS) that had caused major global health concerns.They manifest a range of symptoms ranging from a harmless cold to lifethreatening Pneumonia.Corona pneumonia viruses may cause viral pneumonia complicated by bacterial pneumonia.It may also cause bronchitis, whether viral bronchitis is complicated with bacterial bronchitis.The alarming spread of coronavirus to the human population has triggered a lot of research on detection and mitigation (Lau et al. 2011;Bidokhti et al. 2013;Woo et al. 2010;Shalaby et al. 2021;Saad et al. 2022;Ates et al. 2021;Oh et al. 2018;Wertheim et al. 2013;Tajima 1970).
Video-based monitoring systems are now a common fixture in ports.Originally designed for security, these camera systems have also been used for medical purposes in the past (Le and Jain 2009).These systems are now ubiquitous, from high-security facilities down to the humble corner convenience store.Furthermore, technology has advanced significantly.These cameras are now capable of recording beyond the visual light spectrum, i.e., infrared and ultraviolet (Wu et al. 2014;Natta et al. 2020).
Earlier cameras were low resolution and are limited to the visual spectrum (from Red to Violet).Advancements in technology made Infrared and Ultraviolet detection a commonplace.Infrared emitted by a body is actually directly proportionate to its temperature.Thus, detecting the Infrared emission of the body is a good way of detecting body temperature (the same technology used in Night Vision).Thermal cameras are sensitive to any environmental temperature variations.Hence, it can provide more details than the conventional camera.Unlike conventional cameras, the thermal camera is robust enough against harmful weather conditions and wider light spectral variations, e.g., harmful UV.Coupled with sophisticated algorithms such as deep learning, these cameras can function as image recognition systems (Kris ˇto and Ivas ˇic ´-Kos 2018).
The proponents cannot stress enough the need to use all possible technology advancements in mitigating social problems such as pandemics and security threats.Camera or video technology is one of them.The ever-broadening spectrum and ubiquity of camera technology should also broaden its purpose of use.They should be used for managing a local village's nighttime security to contain the spread of a new virus.To support this point, in Lin et al. (2022); Kong et al. 2005;Gyaourova et al. 2004;Singha et al. 2004;Peng et al. 2016) the authors prove that the thermal face recognition system is more optimal than the conventional video-based security systems.
In this paper, automated, deep learning-based, mass pedestrian fever detection is proposed.The proposed system uses deep learning on pedestrian thermal data sets collected using thermal detection capable camera systems.Since Infrared is the target spectrum, this system can be used at night as well as daytime.Furthermore, a conventional camera is utilized to determine the target's identity.Identity is established doing cross-referencing thermal and conventional camera data and existing Big Data.From simulation results, the proposed system can achieve promising performance in mass fever detection.
This paper is organized as follows: ''Pedestrians detection using deep learning based on thermal imaging'' section introduces the thermal spectrum and the pedestrian detection based on neural networks; the proposed system architecture and its operation are described in ''The proposed system architecture'' section; ''The mathematical model'' section is where a statistical model is developed on patients at different detection probabilities; the simulation results are explained in ''Result and discussion'' section and finally, ''Conclusions'' section concludes the paper.
The images in the NIR and the SWIR are affected by the daylight (noise band).Therefore, these bands are used for indoor applications.While MWIR and the LWIR bands, are mostly used in outdoor facial recognition applications.This is evidently because the latter two bands are significantly higher than the noise band (daylight band).However, the images detected in these bands are highly affected by any environmental temperature variation.Additionally, eyeglasses distort this IR spectrum (Kong et al. 2005).Consequently, faces with eyeglasses will be veiled.To overcome these challenges, thermal and conventional camera imaging fusion techniques are used (Gyaourova et al. 2004;Singha et al. 2004).
Many approaches to thermal human recognition is through convolutional neural network (CNN) architecture (Peng et al. 2016;Wu et al. 2016;Sarfraz and Stiefelhagen 2015;Palm 2012;Kwa ´sniewska et al. 2017).CNN technique has relatively higher accuracy rates than traditional techniques of image processing features extraction.The main structure of the CNN is shown in Fig. 2.
The used input datasets differ according to the application.Some researchers have used their own dataset.While others have utilized some benchmark datasets.Then, the features are extracted from the images using the convolutional layers associated with its activation functions.Each convolution layer is a dot product between the features filter coefficients and the input image.Hence, activation function such as Rectifier Linear Unit (ReLU) is applied to the filter output.Thereafter, pooling layer (such as Max-Pooling) is usually performed after each convolution layer has reduced the weights and the dimensions.At this stage, the important features are extracted.Finally, the features are prepared for the fully connected layers, which are basically traditional neural networks.Consequently, the Different metrics can be utilized to evaluate the performance of CNN (Fu et al. 2017).For the classification problem, the confusion matrix is the most important tool for investigating the CNN performance.Additionally, some other common metrics are available such as Accuracy, Precision, Recall, Specificity, and F 1 score.The accuracy is defined as the proportion of corrected predictions over all cases as, where T P ; T N ; F P and F N are the true positive, the true negative, the false positive and the false negative values.The precision can be defined as the ratio of the T P to the sum of all the positive detections as follows, On the other hand, the Recall metric can be used to determine the detection rate as the percentage of the correctly classified positive class.Hence, the Recall and the Miss Rate can be written as, While, the Specificity metric is used to calculate the proportion of the correctly classified negative class.It can be written as, The proposed system architecture Figure 3 illustrates how the proposed system works for monitoring and tracking people possibly infected with the virus.It should be noted that the system works only at the time of a pandemic for resource conservation purposes.In this system, once the conventional camera takes a picture of the person to cross-reference his identity from the Big Data system, the thermal camera takes an IR image of the same person.These cameras are to be installed in well planned strategic locations such as carpark poles, entrances etc.One recommendation is to install the cameras at poles 100 m high in congested traffic areas and 500 m in low traffic ones, so that they can capture the face and the throat.The thermal images of the people are collected individually for better accuracy.All of the thermal cameras are set at certain threshold temperature.This temperature should evidently be higher than the healthy human body temperature, i.e., temperature medically defined as fever.
To confirm if the person is infected, an algorithm is used to cross-reference the collected data from the Big Data for identification.A text message is sent on his cell phone to guide him to the nearest CT station in the same location.The same person goes through further test protocols and triage established for pandemic mitigation.
Expectedly, in a pandemic, a large number of suspected infections will be detected and collected.The proposed system is part of a bigger pandemic mitigation system such that suspected infections are automatically scheduled for isolation and assignment to a CT station.Upon confirmation of positive COVID-19 case, the system automatically arranges proper pandemic transportation at the patient's location.A mobile app with data and Bluetooth capability will provide the patient some information and protocol on how the treatment and isolation will proceed.The system is constantly updating the provided transportation the exact location of the patient or suspected infection.In rare cases where the patient has no access to the app, the CT station personnel are also informed and advised to orient the Fig. 2 The convolutional neural network (CNN) structure (Peng et al. 2016) patient on the isolation and treatment protocol.A basic illustration on how the system work is seen in Fig. 4.

The mathematical model
The mathematical model for the proposed system based on three main factors: (a) The announced daily number of infected people by WHO in a certain country.This will be expressed as the function, yðxÞ as x is the number of patients per day.
(b) Since yðxÞ is not exact and cannot predict the number of infected people in the country under consideration, another factor will be introduced to consider the variability in the number of infected people -f x ð Þ.This function represents the probable number of infected people as a Gaussian distribution, which is most likely to the studied case.This distribution depends on its mean, l and variance,r 2 x .These two parameters will be estimated using yðxÞ especially at the beginning of virus spread.123 (c) The probability of detection, P d which represents the percentage of detection of the proposed system over its coverage area.This percentage depends on the distribution of the smart devices over this area, the technical specifications of these devices, weather conditions, etc. (U.S. Army Night Vision and Electronic Sensors Directorate 2001; Ratches 1999; Keßler et al. 2017;Ratches et al. 1975;Vollmerhausen and Jacobs 2004 ).
In order to estimate the mean, l and variance,r 2 x of the given Gaussian distribution in (1), let us consider the function yðxÞ as a sample represents the population of the actual patients, consider also that m is the Median, a is the minimum value, b is the maximum value and n is the size of this sample (Moyer 2006;Mood et al. Chan et al. 2004;Mastro et al. 1997;Ates et al. 2022).The sample elements will be assorted in an ascending way and then can be expressed as: where the Median is the Mth number, then m ¼ x M if n is odd; where M ¼ nþ1 2 , and To estimate the mean l, the following inequalities can be used.First, for the data set, fx 1 ; x 2 ; . ..; x M g a x 1 ¼ a a a x 2 m a x 3 m . . .
Second, for the next data set, fx Mþ1; x Mþ2 ; . ..; The sample mean, l is the middle column after adding up and diving by n.Also, for all the three columns, adding up and diving by n can drive the following inequality: substituting (2) in ( 5) therefore, the lower and upper bounds of the mean can be defined as: The upper Bound Then, the estimated mean can be expressed as: Therefore, the sum of mean squares can be estimated as: Fig. 8 The system performance evaluation Using the computational formula, the sample variance can be calculated as: However, we can estimate P n i¼1 x 2 i and 1 n P n i¼1 x i , then using ( 8) and ( 9), the sample variance can be expressed as: Finally, the estimated number of patients can be expressed by y est ðxÞ given by:

Pedestrian and patient detection
In this proposed system, the thermal camera can measure pedestrian temperatures to detect fever.While a conventional camera provides information that is more detailed on the pedestrian's face features.Pedestrian initial detection is done using the thermal camera.That is due to its ability to work in both nighttime and daytime.If the pedestrian is detected, his temperature is evaluated to determine if the target is feverish.
The used dataset for training the proposed model is collected from several available thermal camera videos for detection of pedestrians approaching or waiting at the crosswalk (Kris ˇto and Ivas ˇic ´-Kos 2018).Samples of the video frames are shown in Fig. 5.The sample video has 29 frames/sec with frame dimension of 368 9 288 pixels.The corresponding pedestrian distribution with the time of the sample video is shown in Fig. 6.
For training, a reduced FLIR thermal dataset is used (which can be free downloaded at http://www.flir.com/oem/adas-dataset-form/).Only one class is selected (pedestrian) with 13,094 samples.While 1247 samples are used for testing.These samples (as shown in Fig. 5) are collected from the video frames accessed by the following link: http://www.youtube.com/watch?v=yODTcCwOcV4).The pedestrian detection is based on the aggregate channel features (ACF) algorithm (Dollar et al. 2014).It converts the channel of the extracted region into LUV color space channel.Then, the gradients magnitudes are computed for the Histogram of Gradient (HoG) calculation.Further, the features channels are aggregated and convolved to extract the feature vector.Eventually, it is arranged in a boosted The performance of the pedestrian detection system can be measured using the precision and the miss rate metrics  as shown in Fig. 8. From results, the system can achieve 0.9 average precision and 0.3 average miss rate.

Image registration and face recognition
Image registration is an elimination process within a full motion scene.Both fixed and moving images are used in producing one coordinate domain.Figure 9 shows an example of a successful registration.Figure 9a displays two overlapped images before the registration process.While, Fig. 9b shows the overlapped images of the same subject after the registration.It can be shown that the image registration reduces the shifted pixels relative to both frames.This increases the signal to noise ratio of the resultant image (Kola ´r ˇova ´and Bernard 2015).
The proposed system performs pedestrian detection and registration as described earlier.The outputs of these processes are then used to improve or train the system.Thus, fusion between both the thermal and the normal images increases the accuracy of the captured patient image.After the registration process, the face recognition technique is performed with the help of the Big Data center to recognize the patient's identity.

Virus spreading detection
This system is primarily for detection, collection, Neural Network training and early prediction of deadly viral infection to mitigate epidemic and pandemic spread of a virus.Therefore, the proposed system will provide a learning (Neural Network) that could be deployed in the early onset of a deadly viral spread.Further, this system is capable of learning from Big Data established from previous pandemics that could be used for future pandemic mitigations.The greater the likelihood infections are detected; the less likely epidemics will occur.In these results, the effect of three values of the probability of discovering the infections was presented based on three real examples.These results were announced by the WHO from Feb., 15 to Apr. 8, 2020 (https://covid19.who.int/).
Figure 10 shows the numbers of infected people due to the spread of Corona virus in three countries of the world USA, Italy and Saudi Arabia in the mentioned periodic data according to WHO.From Feb., 22 to Mar.25, Italy ranked first globally in the number of people infected with Coronavirus, but the situation changed after that, and USA became the first in the world by a large difference in the number of infected people.
Although a country like Saudi Arabia at the end of this period is considered to have no more than 1 percent of those infected in USA.This is due to the rapid application of precautionary rules to limit the spread of infection and also the population density per square kilometer which has a very significant impact.Among these rules is the application of the curfew throughout the day or part of the day.
In Figs.11, 12 and 13, the proposed system has been applied to the three countries mentioned previously, with a probability of detecting 90%, 85% and 80%.With the onset 13 Normal spreading distribution of corona patient in KSA versus the suggested system spreading at different detection probabilities of the spread of the infection, the system is activated and compared to the normal situation.Obviously, it appears from all results that, by increasing the percentage of probability of detection, the number of infected cases increases over the number that appears in the normal case of using precautionary instructions only where the situation is controlled and the results from the use of this system are distinguished in the less-spread of infection and blockade it in less time.

Conclusions
This article proposed a new early detection system for the COVID-19.The main purpose of this system is to limit the spread of this virus and some other diseases which work with the same mechanism.The proposed system depends on the available modern and smart technologies like thermal cameras and Big Data system.In this system, two synchronized pictures of the patient face and his body temperature are taken using a conventional and thermal camera, respectively.The data are sent to the Big Data system to establish identity.Once it shows a higher temperature than a certain threshold value (normal body temperature), the patient receives a cell phone message to have another check and confirmation in the nearest CT satiation, since fever is just a symptom.In the mathematical model of this system, the reported data from WHO is considered as training data set to train the system.Hence, its median, minimum and maximum values are used to estimate the population mean and variance.This leads to improved system performance as the system learns.The figures representing the results are related to the distribution of the smart devices over the coverage area along with their technical specification.This model was applied on the data of three countries; USA, Italy and KSA with different values of the probability of detection.In all these examples, the system shows high efficiency and possibility for limiting the virus spread in a short period.

Fig. 1
Fig.1The heat emission spectrum by the human body(Lin et al. 2022)

Fig. 5
Fig. 5 Samples of the thermal camera dataset

Fig. 9
Fig. 9 Image registration: a Overlapped images before registration.b Overlapped images after registration (Kola ´r ˇova ´and Bernard 2015)

Fig. 10
Fig. 10 Normal spreading distribution of corona patient in three different countries from Feb., 15 up to Apr. 8, 2020 as a case study

Fig. 11
Fig. 11 Normal spreading distribution of corona patient in USA versus the suggested system spreading at different detection probabilities

Fig. 12
Fig. 12 Normal spreading distribution of corona patient in Italy versus the suggested system spreading at different detection probabilities