Object Detection Using YOLO Models

International Research Journal of Engineering and Technology (IRJET) e ISSN: 2395 0056

Volume: 09 Issue: 05 | May 2022 www.irjet.net p ISSN: 2395 0072

Object Detection Using YOLO Models

Abstract The world is currently shifting forward with rapid technological development, innovation and studies. This has helped to enhance human being’s lives that will have a broader perspective. The present day technology in diverse drones and sensors has helped in numerous methods to get the item to meet the desires.

Drones are broadly used ultra modern world for plenty purposes. This includes taking pictures, live data consisting of live surveillance structures utilizedbypolice or infantrymento protect and secure regions under their control. It is consequently used as a surveillance device, this paper focusses on the improvement of the chance detection to ensure a faster more accurate and smaller human intervention model and subsequently proposes a model algorithm to use hazard reporting without the want for police interventionat the scene which makes the system absolutely independent. This will make casualty reporting less difficult and less complicated for the police and the person who is a part of the accident as there may be no need to call the police and there is no necessityto be physically present at the scene of the accident.

In this paper we suggest the usage of You Only Look Once (YOLO) algorithm for human like monitoring of roads from a bird’seyeview.

1. INTRODUCTION

Road Traffic accidents have become very common nowadays.Asmajorityofpeoplearebuyingcarsandother automobiles, the incidences ofroad accidents are simplygrowingeach day. moreover, the roadshave becomenarrower, and thetownshaveturn out to begreatlypopulated.

A total of 2,403 cases of road accidents happened on Expressways that caused injuries to 1,997 persons and deathsof1,389persons.Themostnumberofcausalitiesin road accidents were reported on the National Highways accountingfor34.4%(53,213outof1,54,732)followedby StateHighways(25.6%)(39,624deaths).Altogether60,506 personsdiedduetoroadaccidentsontheotherroadsduring 2019.

A total number of 25 patrolling vehicles are deployed on AgraLucknowexpresswaywhichisastretchof302km.In

manycasesthefirstrespondersareownersandemployeeof roadside eateries which works fine if an accident occurs nearbythesame.Sincemostoftheareasareisolated,this resultsinthelagofresponsetimewhichbeingacrucialfirst steptothesavethelifeofthepeopleinjured.

Thisscenariogivestherisetotheneedtomonitortheroads 24/7 which can be achieved in an efficient manner using UAV’s.

Currently,classicalobjectdetectionstrategiespredicatedon region proposals comprise of region based Convolutional Neural Networks (R CNNs), spatial Pyramid Pooling Networks (SPP net), Fast R CNNs, Faster R CNNs, and Region based Fully Convolutional Networks (R FPN) . However, these approaches were futile in achieve concurrentspeedbecauseoftheexpensiverunningmethods andincompetencyofregionpropositions.

You only Look once(YOLO) is the most in demand object detectionsoftware program usedin severa intelligentvisionapplicationsbecauseofitssimpleuseand high item detection accuracy.Additionally, in latest years, diverseclevervisionstructuresbasedonhighperformance and overall performance inbuilt structuresare beingdeveloped.

Although, the YOLOstill requires high end hardware for booming real time detection. In thispaper, we firstdiscussreal timeobjectdetectionpropertyoftheYOLO models, in AI based systems with embeddedstructureswithresourceconstraints.

Specifically, we tend to entail the issues associated with excessiveprecisionandconvenienceofYOLOandprovide real timeobject detectionserviceby meansofminimizing overallprovidersdelay,thatremainsalimitationofthepure YOLO.

International Research Journal of Engineering and Technology (IRJET) e ISSN: 2395 0056

Volume: 09 Issue: 05 | May 2022 www.irjet.net p ISSN: 2395 0072

usingthesetogivetheaverageprecisionofeachfactor.The comparisonisdonebetweentheaverageprecisionthatthe testsetgivesversustheaverageprecisionofthereal world imagescapturedbythemodel.Theoutcomeshowsabetter objectdetectionaccuracy.

Trafficmonitoringisnotpossiblewhendoneinastaticway so to evolve this method of static watch W. Fang et al. [2] emphasisonobjectdetectionusingYOLO,R CNN,andDPM. TheYOLOv3versionofYOLOhasbeendiscussedalongwith theadvantages,drawbacks,andimprovementsoftheYOLO algorithm.ThecharacteristicsofYOLOalgorithmarethatit islightningfastandquick,globalimagespecificationiseasy, it’seasiertogeneralizeandrepresentimagesetc.

Figure1:ObjectDetectionusingYoloalgorithm

Thus,arisetheneedtoenhancethepre existingmodelsof the YOLOalgorithm. So, we will be comparing the fpsand mAPofthecurrentexistingmodelsofthesame

2. LITERATURE SURVEY

Variousresearchhasbeendoneonobjectdetectionfroma verticalview.Thisviewpresentsmanychallengesofitsown few being smaller objects present in the background, viewpointchanges.

The number of computer vision (CV) tasks like object detection and image segmentation have gained extreme acceptancewithinthepastfewyears.Objectdetection(OD) is difficult and helpful for determining the various visual objects of a particular class (such as cars, pedestrians, animals,terrains,etc.).

Object identification in this aspect has been a subject of interest for computer vision analysis for drone based applicationsandautonomousnavigation.

Though the best results can be seen by using two stage networks (YOLO + R CNN) this results in loss of speed of detecting objects which is not favorable in the real world scenariossincedronenavigationandpredictionshouldbe quick to give accurate time bound result. C. Liu et al. [1] Image processing using YOLO network to enhance the detectionoftrafficsigns.

Theproblemareawhichwashighlightedisnotgoodenough imagesbeingcapturedtheproblemofunderexposure,blur, rotationpersists.Theworkisbasedonrealisticviewsofthe real worldscenarios.Themodelistrainedonthedatasetof actual images to make the model more authentic for real worldimagedetection.TheYOLOneuralnetworkisusedto analyze the object which are the traffic signs, and the analyses was based on Darknet 53 network structure. Furthermore,theanalysisisdoneoncertainfactorsnamely Blur of the image, Flip or rotation, noise and cropping by

Thenetworkdesignhasbeendoneasaconvolutionalneural networkandthedatasetusedisPASCALVOC.Thenetwork has been designed to maximize the extent of fast object detection.Thedataistrainedandthemeasureadequately thedetectoridentifiesthelocationsandcategoriesofobjects throughoutnavigationispredictedwiththehelpoflogistics regression, further the object class is predicted, and the inferenceisgeneratedfortheimage.

The limitations stated in this research comprise of the differenceinresultsgivenbythehugevariationinthesizeof theboundingbox.Inaccuraterestrictiontoaparticularplace is the major setback that has been observed due to the differenceobserved.

ThisworkfurthernarratesthefuturescopeoftheYOLOand YOLOv3 algorithm considering the COCO dataset which concludes the flexibility and accuracy achieved using the YOLOv3overtheformerlyusedalgorithmswhencompared inareasofimagesdetectionandclassification.Whentalking aboutrealtimesurveillance J. Tao et al. [3]talksaboutthe useofcomputervisionforobjectdetectionalongwiththe deep learning modules particularly convolutional neural networkandYOLO.Themotiveofthisresearchistoidentify andlocateobjectinatrafficroute,andtobeusedfurtherfor the purpose of surveillance of the traffic. The comparison between the traditional machine learning algorithms and deep learning algorithms for the prime purpose of object detectionhasbeendoneandimplementationshowsthebest outcomeintheYOLOalgorithm.

Theneedforrealtimesurveillanceandaccurateresultsin different scenarios has led this research to happen. The recentrelatedworkdoneinthisdomainpointedtotheuse ofconvolutionalneuralnetworkasthefarbestapproachin getting the expected results. The factors of the designing schemelikenetworkdesign,combiningOYOLOandR FCN and pre processing the images are elaborated on. The experimentation includes the KITTI dataset and further explainthetrainingprocessandtheoutcomeachieved.

Adarsh et al. [4]explainstheneedofenhancementsinthe detectionspeedandaccuracyhasbeenaprimeimportance

International Research Journal of Engineering and Technology (IRJET) e ISSN: 2395 0056

Volume: 09 Issue: 05 | May 2022 www.irjet.net p ISSN: 2395 0072

whenthemonitoringtrafficandprovidingaccurateresults foranycausalityusingdifferentobjectdetectionmethods, whichareHOG,RCNN,FastRCNN,FasterRCNN,YOLOv1, YOLOv2,YOLOv3,SSD,etc.

The implementation and analysis give varied results for a particular characteristic like speed, accuracy, matching strategy, IOU threshold, training dataset, pace of learning, etc.

UsingYOLOv3tinyincreasesthespeedofobjectdetectionin addition to better accuracy, on static object and on video containingdynamicpictures.

Imagerecognitionisanotherfactorthatfacilitatessmooth workingofthetrafficsoforthat Ratre et al. [5]explainsthe YOLOv3 version of YOLO which has been discussed along withtheadvantages,drawbacks,andimprovementsofthe YOLOalgorithm.ThecharacteristicsofYOLOalgorithmare thatitislightningfastandquick,globalimagespecification iseasy,it’seasiertogeneralizeandrepresentimagesetc.

Thenetworkdesignhasbeendoneasaconvolutionalneural networkandthedatasetusedisPASCALVOC.Thenetwork has been designed to maximize the extent of fast object detection.Thedataistrainedandthemeasurehowwellthe detectoridentifiesthelocationsandclassesofobjectsduring navigationispredictedwiththehelpoflogisticsregression, further the object class is predicted, and the inference is generatedfortheimage.

The limitations stated in this research comprise of the differenceinresultsgivenbythehugevariationinthesizeof theboundingbox.Inaccuraterestrictiontoaparticularplace is the major setback that has been observed due to the differenceobserved.

ThisworkfurthernarratesthefuturescopeoftheYOLOand YOLOv3 algorithm considering the COCO dataset which concludes the flexibility and accuracy achieved using the YOLOv3overtheformerlyusedalgorithmswhencompared inareasofimagesdetectionandclassification.

But the YOLOv3 model conflicts with the FF YOLO model basedonaccuracy, L.Yitong et al. [6]speaksonusingthe machine learning YOLO model for object detection in complex scenes by feature fusion. The work in this paper talks about the backdrop of YOLO V3 model in object detectionandthussupersedeitwithFF YOLOfortheaimof faster and more accurate object detection in complex scenarios

Fortheneededimprovementafour scaledetectionlayerin incorporatedtothealreadyexistingthreescaleprediction mechanismformorepreciseinputforanupgradedoutput.

PascalVOC2007andMSCOCOdatasetareusedinthispaper forcomparisonofmAPondifferenttargets.Thecomparison

has been done between YOLOV3, YOLOV4, and FF YOLO models and the comparison has been done based on the number of targets detected and on those parameters the detectionaccuracyhasbeencalculatedwhichshowedthat FFYOLOgaveresultsbetterthantherestinfuzzyorcomplex imageswithoverlappedboundingboxes.

Xianbao et al. [7]improvisedtheYOLOv3modeltofurther refinetheprocessofimagedetection.

Now we have the superseded version of YOLO algorithm whichifFFYOLObyYOLOv3owingtoitsbetteraccuracyin detectingimagesbutthatitselfdoesnothelpinmonitoring the dynamic movements. A.Sarda et al. [8] takes a go on detecting objects in autonomous driving using YOLO and computer vision which eliminates the chance of mis happeningsthatmighthappenifhumaninterventionexists. ThemodelofYOLOusedisYOLOV4 andthedatasetusedis OIDV4.

Thedetectionrateofdifferentalgorithmsiscomparedlike theCNN,RNN,SVM,KNNwhichhavebeenreplacedbyYOLO because ofitshighaccuracy andfaster results.Themodel whichisbeingtrainedneedstwoprerequisiteswhicharethe most favorable coordinates of the bounding box and the object class. The work in this paper talks about using the YOLO neural network architecture in two of its models namelyYOLOV3andYOLOV4alongwithmentioningabout theupperhandthatYOLOV4modelhasoverothermodels for data increase using synthetic data and other such techniques.

The data has been tested on 3200 images and trained on 8000images.Furthermore,themodulecanbeenhancedby trainingandtestingonmoredataforpreciseresults.

Figure2:Comparisonofdifferentobjectdetectionmodels

International Research Journal of Engineering and Technology (IRJET) e ISSN: 2395 0056

Volume: 09 Issue: 05 | May 2022 www.irjet.net p ISSN: 2395 0072

totally different programming designs to form easy or complicated programs, get faster results and write code nearlyasifspeakinginaveryhumanlanguage.

3.5 Open CV

Figure3:ObjectDetectionandprecision

3. REQUIREMENT ANALYSIS

3.1 Deep Learning

Deep Learning isasubfieldofmachinelearningandapivotal part of artificial intelligence (AI). It is a neural network which attempt to simulate the behavior of human brain empowering it to learn from a large data set. It is an optimizedneuralnetworkwhichshoesbetterresultwithan enhancedaccuracyrate.

3.2 Artificial Intelligence

Artificialintelligence(AI)referstothesimulationofhuman intelligence in machines. The perfect characteristic of artificial intelligence is its ability to rationalize and take actions that have the simplest likelihood of achieving a selectedgoal

Artificial intelligence(AI) istheabilityofa computer, or a droid controlled by a computer to try to do tasks that are done by humans as those need human intelligence and discernment.

3.3 YOLO V4

YOLO,asstated,standsfor You Only Look Once,itisanobject detectionsysteminactualperiodthatacknowledgesvarious objects in an exceedingly single enclosure. Moreover, it identifies objects sooner and more exact than various recognitionsystems.YOLOisafuturisticrecognizerthathas a quicker FPS and is more precise than available detectors.The detector will be trained and used on a standardGPUthatallowswidespreadadoption.Newoptions in YOLOv4 improve accuracy of the classifierand detector andmaybeusedforotherresearchprojects.

3.4 Python

Python is an interpreted, object oriented, high level programminglanguagewithdynamiclinguisticsdeveloped by Guido van Rossum. It had been originally released in 1991.

It is a multiparadigm, all purpose, interpreted, high level programminglanguage.Pythonpermitsprogrammerstouse

OpenCVisthehugeopen sourcelibraryforthecomputer vision, machine learning, and image processing and currentlyitplaysasignificantroleindataprocessingthat is extremely vital in today’ssystems.By using it, one can method pictures and videos to spot objects, faces, or perhapshandwritingofanindividual.Onceintegratedwith various libraries like NumPy, python is capable of processingtheOpenCVarraystructureforanalysis.Tospot imagepatternanditsvariousfeatureswehaveatendency tousevectorareaandperformmathematicaloperationson thesefeatures.

4. CONCLUSION

Object detection for the use case we are trying to solve requirespredictionata real timespeedotherwiseitis no better than manually monitoring drones. YOLO (V4) in particulargivesadetectionspeedof32fpsondeviceswith lowGPSpowerandgoestoalmost150fpsongraphicheavy devices.Thedronesthesedayshaveamaximumcapacityof 4gborgraphiccardsandwiththeintroductionofpowerful yetlightchips(likeApple'sM1chip)itisonlygoingtorise. So that way Yolo is the best algorithm at this time for detectingobjectsforarealtimeusageasobjectdetectionsin drone

But there are still a lot of issues in a real time use case becausemostofthetimesthevehicleorobjectmaybefar awayandwouldhampertheperformanceandaccuracybut with time and variable speed of the drones can be able to solvethisproblemaswell.

REFERENCES

[1] C.Liu,Y.Tao,J.Liang,K.LiandY.Chen,"ObjectDetection Based on YOLO Network," 2018 IEEE 4th Information Technology and Mechatronics Engineering Conference (ITOEC), 2018, pp. 799 803, doi: 10.1109/ITOEC.2018. 8740604.

[2] W.Fang,L.WangandP.Ren,"Tinier YOLO:AReal Time ObjectDetectionMethodforConstrainedEnvironments,"in IEEE Access,vol.8,pp.1935 1944,2020,doi: 10.1109/ACCESS.2019.2961959.

[3]J. Tao,H. Wang, X.Zhang,X.LiandH.Yang,"Anobject detectionsystembasedonYOLOintrafficscene,"20176th InternationalConferenceonComputerScienceandNetwork Technology (ICCSNT), 2017, pp. 315 319, doi: 10.1109/ICCSNT.2017.8343709.

International Research Journal of Engineering and Technology (IRJET) e ISSN: 2395 0056 Volume: 09 Issue: 05 | May 2022 www.irjet.net p ISSN: 2395 0072

[4]Adarsh,P.,Rathi,P.andKumar,M.,2020,March.YOLO v3 Tiny:ObjectDetectionandRecognitionusingonestage improved model. In 2020 6th International Conference on Advanced Computing and Communication Systems (ICACCS) (pp.687 694).IEEE.

[5]Sharma,A.,Singh,A.,Shetty,C.andRatre,S.,2020,June. YOLO(YouOnlyLookOnce)TechnologyandIts’Impactin FieldofObjectDetection.In Proceedings of the International Conference on Recent Advances in Computational Techniques (IC RACT)

[6] C. Baoyuan, L. Yitong and S. Kun, "Research on Object DetectionMethodBasedonFF YOLOforComplexScenes,"in IEEE Access, vol. 9, pp. 127950 127960, 2021, doi: 10.1109/ACCESS.2021.3108398.

[7] Xianbao, C., Guihua, Q., Yu, J. et al. An improved small object detection method based on Yolo V3. Pattern Anal Applic 24, 1347 1355(2021).

[8] A. Sarda, S. Dixit and A. Bhan, "Object Detection for Autonomous Driving using YOLO [You Only Look Once] algorithm," 2021 Third International Conference on IntelligentCommunicationTechnologiesandVirtualMobile Networks (ICICV), 2021, pp. 1370 1374, doi: 10.1109/ICICV50876.2021.9388577.

[9]towardsdatascience.com/yolo you only look once 17f9280a47b0

[10]towardsdatascience.com/yolo v4 or yolo v5 or pp yolo dad8e40f7109

2022, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal