A Hygiene Monitoring System

Brian Cherin Rohan Jinturkarbrncherin@gmail.com rjinturkar@gmail.com

Bhargavi Lanka Daniel McKeon Nivedhitha Sivakumarbhargavilanka29@gmail.com dftmckeon@gmail.com blueflute19@gmail.com

Alex Weiner*alexweiner@alexweiner.com

New Jersey’s Governor’s School of Engineering and TechnologyJuly 27, 2018

*Corresponding Author

Abstract—In the healthcare industry, patients are often ex-posed to harmful pathogens due to a lack of compliance to handhygiene protocol. The vast majority of healthcare professionalsdo not abide by hand hygiene standards established by theWorld Health Organization, facilitating the spread of nosocomial(hospital-acquired) infections. When representatives trained inproper handwashing procedures monitored medical profession-als, there was a significant increase in compliance with properprotocol. Given this correlation between observance and adher-ence, a Hygiene Monitoring System was developed to monitorhandwashing through the application of machine learning. Theembedded system captured, processed, and compared instancesof handwashing to the proper procedure. An implementation ofthis system would encourage healthcare professionals to followthe official protocol denoted by the World Health Organizationand dramatically reduce the likelihood of healthcare–associatedinfections.

I. INTRODUCTION

Hands are the primary pathways of germ transmission, yethealthcare professionals often overlook proper handwashingprotocol [1]. In one study, only 22% of hospital workerscomplied with protocol [2]. Under surveillance, complianceincreased to 52.2%, demonstrating a trend between observationand adherence. This issue has global ramifications: healthcare-associated infections (HCAI) contribute to over 135,000 deathsin developed nations each year [3]. The Hygiene MonitoringSystem developed in this study monitors and identifies properhand hygiene in hospitals to increase adherence to protocol.

II. BACKGROUND

A. Correct Handwashing Procedure

The World Health Organization (WHO) defines handwash-ing as proper when six motions are performed: the simplepress; the roof ; the webs; the links; the thumbs; and themortar and pestle. Pictured in figure 1, these motions shouldbe performed for at least 40-60 seconds. The webs and thumbs

motions involve the right hand acting on the left and vice-versa; for instance, one must scrub both the left and rightthumbs to abide by WHO guidelines [4].

Figure 1. Top left to right: simple press, roof, web; Bottom left to right: link,thumbs, mortar & pestle

The WHO also recommends that healthcare workers fol-low proper handwashing procedure after five key moments.They are defined as 1) before touching a patient, 2) beforeclean/aseptic procedures, 3) after body fluid exposure/risk, 4)after touching a patient, and 5) after touching patient surround-ings. The concept that healthcare workers must consistentlyabide by WHO protocol is essential to the purpose of thissystem [5].

B. Software

C was the computer programming language chosen forthis project. It was developed in 1972 by Dennis Ritchieat Bell Labs, and is closely associated with the Unix op-erating system. It is commonly considered a middle levelprogramming language, which does not refer to its difficulty

1

or programming power, but rather its ability to access thesystem's more basic functions while still supplying higherlevel constructs. Compared to other higher level programminglanguages such as Python, C is more efficient in its processingabilities, allowing for rapid analysis of adherence to protocol.

Linux is a family of open-source, free operating systemsthat can be installed on a variety of electronic devices. Thissystem utilized Ubuntu Linux 12.04, which came pre-installedon the NVIDIA Jetson TX2. The operating system supportsC programming as well as direct command line input in theform of UNIX commands, also referred to running commandsin the terminal.

Bash is a command processor that allows the system toread commands directly from a file. Various Bash scripts wereexecuted in order to install the libraries defined later in thispaper.

C. Hardware

The NVIDIA Jetson TX2 is an embedded system withmachine learning capabilities. Equipped with USB ports anddigital/analog pins, it is capable of connecting to a varietyof sensors and peripheral devices. The NVIDIA Jetson TX2runs on C++ but has C functionality and works with theUbuntu operating system. The included Graphics ProcessingUnit (GPU) supports the machine learning algorithms used inthis study. The system also supports a Camera Serial Interface(CSI) which is an embedded camera system that captures highdefinition images and videos [6].

A GPU, unlike a Central Processing Unit (CPU), has theability to execute or handle thousands of threads simultane-ously. A GPU contains many more processing units (cores)than a CPU, enabling it to process data very quickly. Thisdramatically reduces computation time for running complexalgorithms, increasing efficiency. This practice of using GPUsto perform computations is referred to as general-purposecomputing on graphics processing units [7].

D. Image Processing

This study involved conversions between multiple filetypes. They are defined below for reference:

• MP4, or MPEG-4 Part 14, is a modern multimedia fileformat for storing video and audio in a space and qualityefficient format with widespread compatibility.

• AVI, or Audio Video Interleave, is an older multimediafile format for storing video and audio with less compat-ibility than more modern formats.

• JPEG, named for its development by the Joint Photo-graphic Experts Group, is a file format for storing imagesin smaller sizes than other formats. It utilizes approxima-tions while compressing pictures and thus cannot supporttransparency.

• PNG, or Portable Network Graphics, is a file formatfor storing images with greater quality than most imageformats as a result of utilizing lossless compression.

Unlike JPEG, it supports transparency but sacrifices asmaller file size for greater quality.

The red, green, and blue (RGB) color model is a format forproducing a broad array of colors by adding varying amountsof red, green, and blue pigment, ranging from the values 0to 255 for each pigment. By describing colors in this format,electronic systems can differentiate between shades. In thisproject, RGB pixel values were extracted from images tofacilitate the identification of hand motions within each image.

The Gaussian Blur uses a Gaussian distribution to blur animage. This method of blurring the image reduces image noiseand increases the accuracy of the Support Vector Machinesused in the system. In this system, two constants necessaryto execute the Gaussian blur, the radius and the sigma value,were assigned to be 10 and 5, respectively.

ImageMagick is a photo and video editing library com-patible with UNIX terminal commands. Its implementationallowed for the elimination of complex algorithms. Throughthe application of a Gaussian Blur, alteration of light contrast,and cropping of images, this library aided in maintaining thevisual consistency of the images throughout this study [8].

FFmpeg is a library used to edit and manipulate informationstored in various video formats. FFmpeg commands availablewithin the UNIX terminal could separate a recorded video intoindividual frames without compromising quality.

Lossless image compression is a format for image datastorage that utilizes additional space in order to preserve imagequality. This image formatting technique was utilized in placeof lossy compression, which would reduce quality in order toreduce the amount of space the image uses in storage. Thelossless image compression provided by the PNG file formatwas chosen because image quality was preferred over storagepreservation while training a machine learning model.

E. Relevant Machine Learning

Machine learning is a subset of artificial intelligence thatprovides systems with the ability to learn and refine modelsthrough experience, as opposed to relying on explicit coding.This field intends to help systems learn automatically tomake better decisions in the future based on given data andexamples.

The K-Means segmentation algorithm, one processing tech-nique common in machine learning, can be used to isolate thehands from the background of the images. K-Means segmentsan image into k groups based on the similarity of pixel colorvalues. If an image contains someone's hands in front of awhite background (i.e. a sink), K-Means with a clustering ofk=2 will classify the pixels into either the group of backgroundpixels or the group of hand pixels. Once the two regions havebeen identified, the relevant pixels can be preserved, while theunwanted background pixels can be removed before the imagedata is sent to the Support Vector Machine [9].

Support Vector Machines (SVM) are learning algorithmsthat analyze data for regression and classification purposes.An SVM attempts to draw a line between two data sets, asdepicted in Figure 2, in order to classify those above the line

2

into one set, and those below the line into another. SVMlight

is an implementation of a Support Vector Machine in theprogramming language C that provides various algorithms.SVMlight's pattern recognition function trained the system toclassify new datasets of hand motions against a standard andcheck for a match.

Figure 2. Support Vector Machine Separation

SVMlight uses kernel functions in order to classify datain different forms. Kernels are methods of organizing data,based on an equation. SVMlight offers the option to create akernel in addition to four predefined types of kernels: Linear,Polynomial, Radial Basis Function (RBF), and Sigmoid Tanh.This system employed linear kernels [K(X,Y ) = XTY ],which attempted to match images to known handwashingmotions based on a linear equation [10].

A trained model is generated by giving a machine learningalgorithm examples of categorized data in order to “teach” itto categorize new information, which it does through kernels.This trained model is what is used to match new images toalready known handwashing images. In SVMlight, a modelis trained with raw data provided in specialized text files.These data files are in a specific format with attributes calledfeatures and values, which are paired together. When new datais collected, it can be compared against a trained model todetermine if the data matches the data represented within themodel.

III. EXPERIMENTAL PROCEDURE

The primary subsets of the experimental process were datacollection, image processing, and machine learning.

A. Preface

The final implementation of the Hygiene Monitoring Systemwould consist of activation, recording, image processing, andhandwashing procedure verification. However, before imple-menting the system, it was necessary to develop a trainedmodel by manually collecting data in a video format beforebeing split into images stored in a PNG format. This trainingdata was then prepared by applying a contrast, a GaussianBlur, and was resized. The data was then iterated through

with a C script to convert the images into raw data to feedinto the SVMlight. When a recording is started, data wouldbe automatically collected, processed, and analyzed by thesystem.

The terminal command nvgstcapture, pre-installed on theJetson TX2, begins the camera recording. This study utilizedthe embedded CSI camera system with a resolution of 1920pixels by 1080 pixels and a frame rate of 60 frames per second.The above command produced JPEG files at an interval ofone picture per second for 40 seconds, for a total of 40pictures. This differs from the collection of training data asthe tremendous amount of data stored in a video is onlynecessary for training the model. Using ImageMagick, theseimages designed for testing were cropped and modified witha slight contrast and Gaussian Blur. Subsequently, a K-Meansalgorithm was applied to the each image to remove extraneousinformation by setting the values of all background pixelsto black. At the end of this process, the images containeda slightly blurred pair of hands with a black background. Thepixel values of these images were then run through a script inC (Appendix Reference to Code) to format the image data tomake them readable by the SVM. This formatting was done sothat the SVM could classify the testing data more accurately,because the images would be as similar as possible to the datathe SVM used to learn. The formatted data was inputted intoSVMlight to test the model.

The model was eventually tested by evaluating the accuracyof the system in identifying a single motion. After performingthe first test, there was more test data created to test the abilityof the system to identify all of the motions correctly whenprovided with an entirely proper test dataset. These varioustests were performed in order to gauge the efficacy of thetrained model in various expected circumstances.

B. Segmenting Training Data into Images

The collection of data was necessary in order to train themodel. Video was recorded of the subject performing thesix handwashing motions, centered in a well-lit environmentwith a white background. The demonstration was performedwithin two feet of the lens. The videos were recorded andsaved in an AVI video format. The data consisted of thehandwashing motions each recorded for approximately 55seconds in the order of R over L web, L over R web, Link,R Thumb, L Thumb, Simple Press, Roof, and Mortar Pestle,where “R” represents the right hand, and “L” stands for theleft hand.

Afterward, the video was edited to remove transitionsbetween motions and rendered into an AVI format beforebeing split into individual images in a PNG format. Throughthis process, the eight minutes and six seconds of video wereseparated into 28,806 individual images using the FFmpeglibrary in the UNIX terminal. The command

ffmpeg -i 5motions scoop.avi ./Frames/motions%05d.png

was run, where 5motions scoop.avi was the rendered video

3

of purely the hand motions and ./Frames/motions%05d.pngsaved each individual frame of the video into a directorynamed Frames.

C. Image Processing

1) Preprocessing of Training ImagesImageMagick was used to alter the images before using

them as training data. The command

mogrify -shave 300x75 -contrast -brightness-contrast5 -gaussian-blur 10x5 “filename”

allowed for the cropping of 300 pixels off the sidesand 75 pixels off the top and bottom, the addition of minorcontrast, and the application of a Gaussian Blur. The croppingremoved any undesired details while the contrast intensifiedthe difference between the lighter and darker elements of theimage [10]. This command was added into a bash script thatiterated through each of the 28,806 PNG files in a specifieddirectory to crop and apply a contrast to each image.

A script was then written in C to first apply a K-Meansalgorithm and then to transition each of the PNG files intodata files in a format readable by the SVMlight library. TheK-Means algorithm categorized each pixel of the image intoeither a background class or a hands class. Each pixel assignedto the background class was then effectively removed from theimage by being set to black, while the pixels assigned to thehands class were maintained to be sent to the SVM trainingprocess.

The reformatting algorithm proceeded to break down eachimage into a text file consisting of many rows representing thepixels of each image. Each row consisted of attributes calledfeatures and values in the format of feature:value, where thefeature was the index of every fifth pixel and the value wasa representation of the color of that pixel. An interval of fivepixels was chosen to significantly reduce the file size requiredto store the data. The color value used to represent each feature(pixel) was the sum of the red value multiplied by 100, thegreen value multiplied by 10, and the blue value multipliedby 1. In this way, three values of color (red, green, blue) wereconsolidated into one large value to match to every fifth pixel.Together, these rows of features and corresponding values werecombined into text files that would be interpretable by theSupport Vector Machines.

2) Testing Images’ Processing PipelineIn order to accurately test data against the trained Support

Vector Machines, the images supplied to the machine learningalgorithms needed to appear in a similar format to thoseused to train the model. Thus, the image adjustments andbackground removal from the preprocessing steps were alsoused in the image processing pipeline.

The images collected from the Jetson TX2’s camera werealso resized to match the training images. Each image wasthen provided with a level of contrast, brightness, and aGaussian Blur after being resized using the ImageMagickcommand

mogrify -resize 1320x930\! -contrast -brightness-contrast5 -gaussian-blur 10x5.

These images were then converted into the PNG formatwith the ImageMagick command

magick convert filename.jpg filename.png

where filename.jpg is a JPEG file captured from theCSI camera on the NVIDIA Jetson TX2 and filename.pngrepresents the output of the new file.

Next, to remove any bias that would be introduced by irrel-evant pixel data, the background of each image was removedto isolate the hands in the image. This was an essential step,as isolating the hands generated a clearly-defined shape andoutline for each hand motion. This information, in the formof the absence of the surrounding pixels, would be a signif-icant factor in the machine learning process. The K-Meansclustering algorithm was used to remove the backgrounds.This algorithm segments an image into k groups based on thesimilarity of pixel color values. The algorithm treats the datain this system as a three-dimensional system, with dimensionsof red, green, and blue color values. This allowed for thedistance between the color values of two given pixels to benumerically calculated, thus providing a direct measure of thepixels'similarity.

The general K-means algorithm proceeds as follows:1) Select k initial centroids (the center of each group)

randomly within the bounds of the set of data.2) For each data point, find the closest centroid to that

point by calculating distance values between each pixeland the centroid pixels. Assign the point to the groupbelonging to that centroid.

3) For each group, take the RGB values of each pixel andadd them separately, so that there is a total red value, atotal green value, and a total blue value for all the pixelsin that group.

4) Divide each value by the total number of pixels in thatgroup in order to get the average red, green, and bluevalues.

5) Set the centroid pixel’s color value of that group equalto those average RGB values.

6) Repeat steps 2, 3, 4, and 5, until the location of eachcentroid does not change.

For the purposes of this system of SVMs, a value of k = 2was established in order to segment the image into two parts:the background and the hands. However, the initial resultsfrom the K-Means algorithm appeared to be inaccurate andinconsistent, resulting in unsuitable training data. This waslikely because of the random initialization of the centroids,which may have resulted in color values that were too closeto each other to allow for effective separation of the imageinto groups.

To accommodate for this, rather than beginning with randomcentroids, a targeted centroid was selected as the average color

4

of the top-left corner of the image (an area assumed to be partof the background), and a second centroid from the averagecolor of the center of the image (an area assumed to be thehand). At the end of the algorithm, the group resulting from thecentroid from the top-left would contain all of the backgroundpixels, and the group resulting from the centroid from thecenter would contain all of the relevant hand pixels. Once thegroups were determined, the pixels in the background groupwere changed to black pixels. This effectively removed thebackground pixels from the machine learning training process,isolating the hands from the background of each image.

The final product of these image processing techniques wasan image of hands that retained their most distinguishingcharacteristics surrounded by a black background. Figure 3depicts the results of each image processing technique.

Figure 3. Left to right: example hands image, Gaussian blur applied, K-Meansapplied to isolate hands

At this point, the algorithm written in the C programminglanguage used in the preprocessing step was again used toconvert the images into the text files readable by the SVMs.The data was then compared to the trained model created viathe SVMlight library within the Ubuntu 12.04 command-line.(Appendix Reference to Code)

D. Machine Learning

Machine learning was used to analyze processed images andcreate a trained model representing the data. The trained modellater served as a reference to categorize given hand motionsas correct or incorrect.1) Using Data to Train a Model

A process that can identify a certain hand motion as propermust be formatted to have features identical to those of thetrained model. Once the image data was formatted correctly,the learn function of SVMlight was used to create a model(support vector machine) for each of the eight hand motionvariations. To do so, correct training examples of each handmotion were fed into the machine learning algorithms ofSVMlight. The models were created while retaining the greatestamount of detail using a lossless compression format withPNG image files and splitting the original video into 60 framesper second.

The data was then converted into a format easily interpretedby the SVMlight library and its algorithms by using a C script(Appendix Reference to Code). Once converted into a suitableformat, the data was used to train the model with the command

svm learn training data model,

where training data was a file containing rows of image

data to train on, as described previously, and model was theoutput file name for the resultant trained model.

2) Comparing Collected Data to the ModelThe final step in the implementation of the Hygiene

Monitoring System was to classify recorded data as correct orincorrect. Classification occurred through the comparison ofcollected data to the trained model, using SVMlight command

svm classify collected data model results

where collected data was a file containing the datarecorded, which would be compared against the model file,containing the trained model, and results was the outputof the results of the comparison. The output file containedinformation that described how close the collected dataadhered to the model by reporting the percentage of the sixhandwashing motions completed out of those representedwithin the model to determine the person’s overall conformityto the procedure. A value of 80-90% would demonstrate thecompletion of a correct handwashing procedure. A value of100% is not expected because several frames of a sampletest data set are expected to contain no hand motions, suchas the transition between two hand motions. However, if apercentage significantly lower than 80-90% was returned, itwould be clear that incomplete handwashing was observed.

3) Testing the ModelTwo tests were performed to evaluate the accuracy of

the Hygiene Monitoring System. The first test attempted tocompare data comprised of the link motion against the trainedmodel. This returned a percentage of compliance to properhandwashing procedure, depending on the motions detected.Then, the percentage returned was analyzed to determine ifthe link motion was detected.

In the second test, new testing data was recorded andprocessed in an identical format to the training model data inorder to perform more intensive testing. Thus, the new imageswere processed with a Gaussian Blur, K-Means partitioning,resizing to 1320 pixels by 930 pixels, and an applicationof light contrast. This new data included the simple press,webs, mortar and pestle, thumbs, and roof motions, omittingthe link motion due to data collection limitations. The resultwas indicated by the percentage derived from the correctlyrecognized images out of the total number of images tested foreach specific motion. Subsequently, the data was also analyzedto report the overall percentage of accuracy for all motions,or the number of correctly classified images out of the totalnumber of tested images.

IV. RESULTS

The performance of the Hygiene Monitoring System wasassessed with test data collected in a simulated environmentwith the intent of verifying the functionality of the system ina controlled setting.

This evaluation tested the system’s accuracy in correctlycategorizing images as distinct hand motions. Multiple sets

5

of test data were created and compared against the trainedmodels.

In order to test the ability of the model to detect a singlemotion at a time, the first test example contained only the linkhand motion being performed. When this data was classifiedthrough the support vector machine’s model, it yielded resultsshowing that it fulfilled 14% of the expected handwashingprocedure. Since the model expects eight distinct motions (sixunique techniques, but two are repeated for each hand), theobserved percentage of 14% is extremely close to the expectedvalue of 14.28%.

In the next test, the system attempted to classify a dataset containing images of a complete handwashing procedure.However, due to data collection limitations, the link motionwas excluded from this set. The other hand motions andassociated data were unaffected. The images in the test dataset included two to five frames for each motion. This datawas then compared to the trained model through SVMlight

to identify the percentage compliance to the handwashingprocedure by the test subject. The number of frames that wereproperly identified are indicated below in figure 4. 18 imageswere correctly identified out of 24 tested, corresponding to anaggregate of 75% accuracy in correctly identifying the handmotions.

Figure 4. Results

V. CONCLUSIONS

The results produced indicate that the Hygiene MonitoringSystem can correctly identify compliance to proper handwash-ing procedure.

A. Significance

A successful implementation of the Hygiene MonitoringSystem could dramatically reduce the rate of infection anddeaths related to healthcare negligence. On any given day,about one in twenty-five hospital patients contract at least onehealthcare-associated infection [11]. Although handwashingseems like an elementary process, multiple studies illustratethe immense ramifications of a lack of regard for procedure.This system encourages greater compliance, a major steptoward limiting the spread of disease and prioritizing patientwelfare.

B. Applications

In the future, the Hygiene Monitoring System can be im-plemented in hospitals around the world. The system's abilityto monitor and identify correct handwashing practices will in-crease transparency within the healthcare industry. Moreover,the systems records could give insight into any liability issuesor legal matters, offering an objective viewpoint into the oftensubjective and complex world of medical lawsuits.

The concept of the Hygiene Monitoring System also lendsitself to the food service industry. The spread of germs fromworkers to food accounts for 89% of outbreaks correspond-ing to food contamination [12]. The implementation of theHygiene Monitoring System in restaurants and food vendorscould advance the state of sanitation in todays world andmitigate the spread of foodborne illness.

Furthermore, the system developed in this study demon-strates a computer’s ability to differentiate between humanhand motions. This would prove useful in the interpretationof sign language hand positions, as a system trained withenough images could develop models representing variouswords and phrases. Such a development would significantlyimprove accessibility for hearing-impaired individuals andthose attempting to communicate with them.

C. Future Improvements

Certain improvements can be made to refine the reliabilityof this Hygiene Monitoring System. Most significantly, thecollection and usage of more data will improve the trainingmodels, increasing the probability that real world data will becorrectly classified by the system.

The K-Means algorithm could be improved to suit a widerrange of image cases. The current algorithm operates underthe condition that the hands are in the center of the image.Such a change would

A stereographic camera (a camera that records from twoslightly different positions simultaneously in order to generatea 3D representation of an image) would provide depth to thedata collected. By deriving SVM values from the physicallocation of a hand rather than from the color values of pixels,the system could be more reliable when testing data with handsof varied skin tones, or with a background of a similar colorto the hands.

The Hygiene Monitoring System could be more user-friendly. Currently, the system is activated by a keyboardpress, which is not appropriate for a real-life context suchas a hospital environment. In the future, the system willideally be activated by a motion sensor to detect when anemployee begins washing their hands. Further integration witha hospital system would also be an ideal addition. Eachemployee, before washing their hands, would check into thesystem using an radio-frequency identification device (RFID)tag, and once the employee washed their hands and the systemdetermined the level of compliance with the six hand motions,the system would record the level of compliance in a databaseof employee data.

6

The kernel currently being used in the SVM is linear, butthere are other, more complex kernels that the SVM coulduse as well. The polynomial kernel separates the result of aclassification based on a defined, but arbitrary order. The RBFkernel separates the result of a classification using normalcurves around the data points and sums them so that thedecision boundary can be defined by a type of geometriccondition (topology). The Sigmoid Tanh kernel separates theresult of a classification using a logistic function to definecurves depending on the logistic value being greater than thevalue generated through the model based upon probability.Working with any of these kernels in the SVM could increasethe system’s accuracy.

APPENDIX

ACKNOWLEDGMENT

The authors of this paper gratefully thank the following:project mentor Alex Weiner for his extensive experience andhands-on involvement; Residential Teaching Assistant SiddhiShah for her invaluable support and efforts in coordinationand communication; Research Coordinator Brian Lai and HeadCounselor Nicholas Ferraro for their guidance and feedback;Program Director Ilene Rosen, Ed. D. and Associate ProgramDirector Jean Patrick Antoine, without whom this projectwould not have been possible. Finally, the authors appreciatethe generosity of the sponsors of the New Jersey Governor'sSchool of Engineering and Technology: Rutgers University,Rutgers School of Engineering, The State of New Jersey,

Lockheed Martin, Silver Line, Rubiks, other Corporate Spon-sors, and NJ GSET Alumni.

REFERENCES

[1] “WHO Guidelines on Hand Hygiene in HealthCare”, Apps.who.int, 2018. [Online]. Available:http://apps.who.int/iris/bitstream/handle/10665/44102/9789241597906eng.pdf?sequence=1. [Accessed: 22- Jul- 2018].

[2] Sebille V, Chevret S, Valleron AJ. Modeling the spread of resistantnosocomial pathogens in an intensive-care unit. Infection Control andHospital Epidemiology, 1997, 18:8492.

[3] Klevens R et al. Estimating health care-associated infections and deathsin U.S. hospitals, 2002. Public Health Report, 2007, 122:160166.

[4] “How to Handwash?”, Who.int, 2018. [Online]. Available:http://www.who.int/gpsc/tools/HAND WASHING.pdf. [Accessed:22- Jul- 2018].

[5] “Your 5 Moments for Hand Hygiene”, Who.int, 2018. [Online]. Avail-able: http://www.who.int/gpsc/tools/5momentsHandHygiene A3.pdf.[Accessed: 22- Jul- 2018].

[6] “Jetson TX2 Module”, NVIDIA Developer, 2018. [Online]. Available:https://developer.nvidia.com/embedded/buy/jetson-tx2. [Accessed: 22-Jul- 2018].

[7] “GPU Programming—NVIDIA,” NVIDIA Developer, 2018. [online].Available: http://www.nvidia.in/object/gpu-programming-in.html [Ac-cessed 26 Jul. 2018].

[8] I. LLC, “Convert, Edit, Or Compose Bitmap Images @ImageMagick”, Imagemagick.org, 2018. [Online]. Available:https://www.imagemagick.org/script/index.php#features. [Accessed:22- Jul- 2018].

[9] Ng, A. and Piech, C. “K Means”, 2013. [Online] Available at:http://stanford.edu/ cpiech/cs221/handouts/kmeans.html [Accessed 26Jul. 2018].

[10] “SVM-Light Support Vector Machine”, Svmlight.joachims.org, 2018.[Online]. Available: http://svmlight.joachims.org/. [Accessed: 22- Jul-2018].

7

[11] “Hand Hygiene in Healthcare Settings”, cdc.gov,2018.[Online]. Avail-able: https://www.cdc.gov/handhygiene/index.html [Accessed 22 Jul.2018].

[12] “Food Worker Handwashing and RestaurantFactors”, Cdc.gov, 2018. [Online]. Available:https://www.cdc.gov/nceh/ehs/ehsnet/plain language/food-worker-handwashing-restaurant-factors.htm. [Accessed: 26- Jul- 2018].

8