Person Detection in Maritime Search And Rescue Operations

2,3,4

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

Volume: 09 Issue: 06 | Jun 2022 www.irjet.net p ISSN:2395 0072

Person Detection in Maritime Search And Rescue Operations

***

Abstract There is a growing necessity to arrange a search and rescue operation (SAR) to offer help and health care to the injured due to an increase in the number of persons who participate in various adrenaline activities or adventure tourism and stay in inaccessible regions. The goal of SAR is to search the broadest feasible area of the territory in the shortest amount of time in order to locate a lost or injured individual. Drones (also known as unmanned aerial vehicles or UAVs) are now widely used .Drones are increasingly being employed in search and rescue missions because they can quickly capture a vast, controlled area. However examining a huge amount of recorded data in detail remains a challenge. Even for an expert, it is not easy to find searched people who are relatively small considering the area where they are , often sheltered by vegetation or merged with the ground and in unusual positions due to falls, injuries, or exhaustion. As a result, automatic detection of people and objects in images/videos captured by drones during these activities is crucial. In SAR operations, the key object is the person, however, recorded from a bird’s eye view, and such recordings are not contained in the large data sets on which these state of the art detectors are trained. To achieve the highest possible accuracy of the detection model, the data set on which the model is trained must have similar conditions to those that appear when testing the model, so it is necessary to train the model with a bird’s eye view data. This study reviews various implementations as well as some predictions for the future of SAR operations.

Key Words: object detector, person detection, search and rescue operations, UAV, image processing , machine learning

1. INTRODUCTION

In 2017, 3,116 migrants were killed in the Mediterranean Sea. Missing Migrants (Missing Migrants, 2018). When it comes to finding victims following shipwrecks, Unmanned Aerial Vehicles (UAVs) and Remotely Piloted Aircraft (RPAs) have a significant advantageoversatellitesurveillancesince

they can explore specific locations utilising real time route planning. Researchers recently investigated computervisionsensors,techniques,andstrategies,witha focus on emergency scenarios that put people in danger (such as fire and flood) or are created by others (i.e.

accidents). UAVs and RPAs are frequently used in search and rescue operations to classify scenes and identify commodities. When compared to prior techniques that reliedonhand craftedfeatureextraction,usingbothsemi supervised and supervised machine learning algorithms for aerial photo categorization and object recognition is a good strategy. Deep convolutional neural networks, such as Faster RCNN , Cascade R CNN , RetinaNet , SSD , and YOLOv3 , have been successful at recognising individuals in photographs of primarily urban scenes in recent years, and achieve even higher accuracy than humans. To achieve such high results, deep network models required to be trained on large data sets as MS COCO, Pascal VOC, and ImageNet. Then, in order to acquire good detection resultsorsignificantimprovementsincertaindomainsnot included in big data sets, such as thermal images of the monitored region, some sports scenarios, and so on, it is requiredtotraindeepnetworksontheimagesetfromthe selected domain .The data set on which the model is trainedmusthavesimilarconditionstothosethatemerge whentestingthemodel,henceitisvitaltotrainthemodel with a bird's eye view data to get the best possible accuracy of the detection model. One suitable dataset is the AFO dataset which contains images from fifty video clips of items floating on the lake surface that were acquiredbytheseveraldrone mountedcamerasthatwere used to build AFO (resolutions ranging from 1280x720 to 3840x2160). Weretrieved3647 photoswith39991items from these films and manually tagged them. The training set(67,4percentofobjects),thetestset(19,12percentof objects), and the validation set were then divided into three portions (13,48 percent of objects). The test set contains selected frames from nine videos that were not utilisedin either thetrainingorvalidationsetstoprevent themodelfromoverfittingtotheavailabledata.

2. MOTIVATION

For almost a century, search and rescue (SAR) operations have essentially been conducted in the same manner, utilising technology such as radar to get a general search area or human aided with the assistance of search dogs. Many additional uses of AI and machine learning across sectors, such as training algorithms to detect abnormalities that the human eye would overlook, might bequiteusefulinSARoperations.Thesetechnologieshave thepotentialtobemoreefficient,cost effective,andtime saving.

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

Volume: 09 Issue: 06 | Jun 2022 www.irjet.net p ISSN:2395 0072

3. RECENT WORK

[1] This research looks at how automated search algorithms and network architecture may be coupled to increase the overall performance of the state of the art Mobi leNetV3 Large and MobileNetV3 Small systems, which are designed for high and low resource use cases, respectively. These models are then tweaked and utilised for tasks like object detection. As explained in this article, the MobileNetV3 Large and Small models were created to offerthenextgenerationofhigh accuracy,efficientneural network models. Apps that use efficient neural networks are becoming increasingly widespread, enabling whole new on device experiences. Higher accuracy and shorter latency are two benefits of improved neural network efficiency.Toachievethisweintroduce

(1)Complementarysearchstrategies.

(2)Newefficientnetworkarchitecture.

[2] This work describes an aerial image based identificationand trackingalgorithmfor rescuingpersons in catastrophe situations, which is supplemented by a semi autonomous reactive control system. The following isalistofcontributions:Acolouranddepth basedhuman detection stage, as well as the usage of a Human Shape Validation Filter A Convolutional Pose Machine provides thepositionsofhumanjointsforthisfiltertoexaminethe humanskeletonformofthediscovereddetectionsinorder to remove false positives. When the point of view is rotated in regard to the target objects, an automated Multi Object Tracking method that is scale, translation, androtationinvariant.Arevolutionarymatchingapproach based on stance and look similarities that may re identify persons who have vanished from the scene for an extended period of time. The unique body attitude and appearance based association technique allows tracking and recognition tasks to be completed in exceedingly difficult aerial sequences. Furthermore, due to the filter's design based on the Convolutional Pose Machine, false positiveshavebeenreducedtozero(CPM).

[3] To aid in search and rescue, surveillance camera systems and unmanned aerial vehicles (UAVs) are deployed. Because a single human cannot monitor numerous surveillance screens at the same time for 24 hours, automatic object detection is critical. Furthermore, the object is frequently too small to be seen on the surveillancescreenbythenakedeye. Toreducedetection timeforsmalltargets,wecombinedpicturesegmentation, enhancement, and convolutional neural networks. We comparedtheauto detectionsystem'sperformancetothat of the human eye. Our technology spotted the target in 8 seconds, however it took the human eye 25 seconds to locate the object. For object recognition, we used an SSD module. Future research could use a variety of SSD networks, including concatenation and element sum of

each layer, to improve accuracy . Using numerous UAVs, wecouldboostsearchandrescueoperationsevenfurther. UsingvideostreamsfromseveralUAVsasonegiantimage, distributed deep learning can be implemented across multipleUAVs.CombiningUAVsensorandimageanalysis processing can aid in the optimization of UAV flight parameters, such as UAV position, energy restrictions, environmentaldangers,anddatasharingconstraints.

[4] In this article, we provide a unique technique for recognising persons in aerial photographs of nonurban territory obtained by an unmanned aerial vehicle (UAV). Inboththeregionsuggestionandclassificationstages,the technique leverages two separate convolutional neural networks. Contextual information is also employed to improvedetectionoutcomesduringtheclassificationstep. Experiment results on the HERIDAL dataset obtained 68.89 percent accuracy and 94.65 percent recall, which is superiortoexistingstate of the artapproachesforhuman identificationundercomparablesituations.

[5] The current research is focused on real time human identification aboard a fully autonomous rescue UAV. The built embedded system was capable of recognising open water swimmers using deep learning techniques. As a consequence, the UAV was able to give precise support in anunsupervisedway,boostingtheoperationalcapabilities offirstresponders.Theproposedsystemisdistinctinthat it use a combination of GNSS technology and computer vision algorithms for both precise human detection and the release of rescue gear. While the suggested rescue system was designed to detect open water swimmers, with a few tweaks, it could also detect humans and provide emergency services to those participating in winter sports activities. Because of the high level of detection and classification accuracy achieved, the suggested approach has endless applications in SAR missionsinavarietyofterrainsandsituations.

[6] Unmanned Aerial Vehicles (UAVs) outfitted with multispectral cameras are used in this study to search for bodiesduringmarinerescuemissions.Inthenorthwestof

Spain, a number of flights in open water settings were conducted, utilising a qualified aquatic rescue dummy in dangerous areas and actual persons when weather conditions permitted. The multispectral pictures were aligned and utilised to train a body identification Convolutional Neural Network. A thorough analysis was conducted to determine the best combination of spectral channels for this purpose. Using 1) complete image, 2) sliding window, and 3) precise localization method, three approaches based on a MobileNet topology were tested. The first technique determines whether an input image contains a body, the second approach employs a sliding window toassign a classto eachsub image,and thethird method use transposed convolutions to provide a binary output with the body pixels highlighted. To align the

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

multispectralcamerachannels,theMobileNetarchitecture was updated in all cases by adding custom layers and preprocessing the data. The precise localization strategy, we conclude, is the most suited method, achieving spatial localization close to 1m with equivalent accuracy to the sliding window. We plan to use this system in real world cooperative missions with ground workers to search wider areas with an autonomous rotorcraft weighing 25 kilogrammes, carrying a payload of 10 kilogrammes, and flying for three hours. The data gathered from these real world missions might potentially be utilised to retrain CNNs with additional photographs to increase their performance. Furthermore, our vision system will be enhanced in the future with additional cameras with varying spectral ranges and an integrated GPU mounted ontheUAV.

[7] Because of their capacity to be deployed rapidly and often and fly at low altitudes, unmanned aerial vehicles (UAVs) have a lot of potential for use in natural resource monitoring methods. Texture measures are often used in pixel based image analysis, but their application in an object based context is less well documented. In this study, we employed subdecimeter UAV data to examine texturemeasuresatvariousscalesinobject basedanalysis with the goal of identifying broad functional categories of plants in arid rangelands. The decision tree was a helpful tool for cutting down viable texture measures for simplicity of computation, and the correlation analysis revealed important insights into the variations in correlationoftexturemeasurepairsovermanyscales.The results show that unmanned aerial vehicles (UAVs) are viable platforms for rangeland monitoring and that the shortcomings of low cost off the shelf digital cameras can be overcome by including texture measures and using object based image analysis, which is well suited to very highresolutionimagery.Ourfindingswillbeincludedinto a procedure for monitoring rangeland with unmanned aircraft.

[8] Over hundreds of object categories and millions of pictures, the ImageNet Large Scale Visual Recognition Challenge establishes a standard for item category categorizationanddetection.From2010untilthepresent, the challenge has been organised yearly, with over fifty colleges participating. This study outlines how this benchmark dataset was created as well as the advancements in item recognition that occurred. We explore the difficulty of accumulating large scale ground truth annotation, highlight major developments in categorical object identification, present a thorough analysis of the current status of large scale picture categorization and object detection, and compare computer vision accuracy to human accuracy. We conclude with lessons learned throughout the course of the five year project, as well as recommendations for futureroutesandenhancements.

Volume: 09 Issue: 06 | Jun 2022 www.irjet.net p ISSN:2395 0072 © 2022, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal | Page

[9] Over hundreds of object categories and millions of pictures, the ImageNet Large Scale Visual Recognition Challenge establishes a standard for item category categorizationanddetection.From2010untilthepresent, the challenge has been organised yearly, with over fifty colleges participating. This study outlines how this benchmark dataset was created as well as the advancements in item recognition that occurred. We discuss the challenges of gathering large scale ground truth annotation, highlight key advances in categorical object recognition, provide a detailed analysis of the currentstateoflarge scaleimageclassificationandobject detection, and compare state of the art computer vision accuracy to human accuracy. We conclude with lessons learned throughout the course of the five year project, as well as recommendations for future routes and enhancements.

[10] We look at the subject of automatically recognising personsinthermal recordsandphotographsinthiswork. The thermal videos are filmed on a meadow with a little woodland, with up to three individuals in different posturesanddistancesfromthecamera.TheCOCOpicture dataset comprises RGB photos of a range of item categories, and YOLO is an object detector that has been pre trained on it. In terms of human detection, the proposedstructureusesthediscriminantpowerofSCand LDAtosurpasstherectangularfeature basedmethod.Our nextaimistoreducecomputercosts.

[11] The challenge of automated human detection in thermal pictures utilising convolutional neural network based models that were initially designed for detection in RGB images is studied in this research. On a dataset of thermal images collected from movies shot at night in clear weather, rain, and fog, at varying ranges and with different forms of movement running, walking, and sneaking the performance of the conventional YOLOv3 model is compared with that of a specially trained model. The tests show excellent results in terms of average precision for all scenarios investigated, as well as a significantincreaseinperformanceforhumandetectionin thermalimagingwithasmalltrainingset.

[12] Unmanned Aerial Vehicles (UAVs) are used to power a number of critical computer vision applications, providing more efficiency and convenience than traditional security cameras with fixed camera angles, sizes, and perspectives.As a result, moving further with related research necessitates the development of an unconstrained UAV benchmark. We build a new UAV benchmark in this study by concentrating on difficult circumstances with new level obstacles. Then, for each work, a thorough quantitative analysis is performed utilising the most recent state of the art algorithms. Experimental results show that existing state of the art algorithms perform considerably worse on our dataset

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

duetonewhurdlesthathaveevolvedinUAV basedactual situations, such as high density, tiny objects, and camera motion. To our knowledge, this is the first time such challengeshavebeenextensivelyexploredinunrestrained scenarios.

[13]Existingcountingmethodsfrequentlyuseregression based methodologies and are unable to properly pinpoint the target objects, making further analysis difficult (e.g., high level understanding and fine grained classification). Furthermore, the majority of previous research has focusedoncountingobjectsinstaticsituationsusingfixed cameras. We're interested in recognising and counting items in such dynamic situations because of the introduction of unmanned flying vehicles (i.e., drones). Unlike traditional region proposal approaches, we use spatial layout information (for example, cars generally park in a predictable pattern) to incorporate spatially regularised constraints into our network to increase localizationaccuracy.Toputourcountingapproachtothe test, we provide a new large scale vehicle parking lot dataset (CARPK) encompassing around 90,000 autos collectedfromseveralparkinglots.Itis,tothebestofour knowledge,thefirstandlargestdroneviewcollectionthat permits item counting and includes bounding box comments.

[14] We present a new aerial video dataset and benchmark for low altitude UAV target tracking in this research, a s well as a photo realistic UAV simulator that may be used with tracking algorithms. On 123 fresh and completelyannotatedHDvideosequencesshotfromalow altitude aerial perspective, our benchmark gives the first evaluationofnumerouscutting edgeandpopulartrackers. Weassesswhichofthecomparedtrackersarebestsuited forUAVmonitoringintermsoftrackingaccuracyandrun time. The simulator may be used to Test tracking algorithmsinreal timesettingsbeforedeployingthem on aUAV"inthefield,"aswellastobuildsyntheticbutphoto realistic tracking datasets With automated ground truth annotations to simply augment existing real world datasets. This lays the groundwork for future advances in accuracy and speed. Our suggested UAV simulator, in conjunction with unique assessment methodologies, allows tracker testing in real world circumstances with livefeedbackbeforetodeployment.

[15]Thepurposeofthisstudyistoinvestigatetheeffectof convolutional network depth on large scale picture recognitionaccuracy.Itsprimarycontributionisadetailed analysis of increasing depth networks utilising an architecturewithextremelytiny(3X3)convolutionfilters, demonstrating that raising the depth to 16 19 weight layers greatly improves performance over prior art arrangements. The depth of ConvNet architecture design is another significant feature that we will address in this study. To accomplish so, we tweak different architecture variables and gradually increase the network's depth by

Volume: 09 Issue: 06 | Jun 2022 www.irjet.net p ISSN:2395 0072 © 2022, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal | Page

adding additional convolutional layers, which is made feasiblebytheuseoftiny(3X3)convolutionfiltersatall levels. The goal of this research is to look at the influence of convolutional network depth on large scale image recognition accuracy. Its main contribution is a thorough investigation of increasing depth networks using an architecture with extremely small (3X3) convolution filters, which demonstrates that extending the depth to 16 19 weight layers significantly increases performance over prior art arrangements. Another important aspect of ConvNet architecture design that we shall investigate in this study is its depth. To accomplish so, we tweak differentarchitecturevariablesandgraduallyincreasethe network'sdepthbyaddingadditionalconvolutionallayers, which is made feasible by the use of tiny ( 3 X 3) convolutionfiltersatalllevels.

[16]It'smorehardertotraindeeperneuralnetworks.For training networks that are far deeper than previously utilised networks, a residual learning technique is proposed. Rather than learning unreferenced functions, we explicitly reformulate the layers as learning residual functions with reference to the layer inputs. We show extensive empirical evidence proving that as network depth increases, residual networks become easier to optimiseandachievehigher accuracy.Deepconvolutional neuralnetworkshavebeencreditedwithseveraladvances in image processing. clasification. End to end multilayer deep networkshavelow,mid,andhigh level featuresand classifiers by default, with the number of stacked layers increasingthe"levels"offeatures(depth).Recentevidence suggeststhatnetworkdepthiscritical,andthetopresults on the difficult ImageNet dataset all use "extremely deep" modelswithdepthsofsixteentothirty.Verydeepmodels have also aided many other non trivial visual recognition problems.

[17] Feature pyramids are a common component of recognition systems for recognising objects at several sizes, however they are computationally and memory costly. For building high level semantic feature maps of various sizes, a top down architecture with lateral linkages is designed. This design is known as the Feature Pyramid Network (FPN). It's a significant improvement above regular feature pyramids. Recognizing things at widely different sizes is difficult in computer vision. The standard approach is based on feature pyramids built on picture pyramids. The scale change of an object is countered by moving its level in the pyramid, making these pyramids scale invariant. This characteristic allows a model to recognise items at a variety of scales by scanning the model over both positions and pyramid levels. We've shown how to create feature pyramids in ConvNets using a clean and straightforward architecture. As a result, it provides a viable solution for feature pyramid research and applications without the requirement to compute image pyramids. Finally, our

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

Volume: 09 Issue: 06 | Jun 2022 www.irjet.net p ISSN:2395 0072

findings imply that, despite deep ConvNets' tremendous representational power and implicit scale variation robustness, it is still necessary to actively handle multiscalechallengesutilisingpyramidrepresentations.

[18] For emergency responders, coordinating many Unmanned Aerial Vehicles (UAVs) to conduct aerial surveysisahugedifficulty.UAVs,inparticular,mustflyat a kilometre scale. while attempting to find casualties as swiftlyaspossible It isnecessarytoassistinthisprocess. it is desirable to take use of the rising availability of disaster data from sources such as manned reconnaissance, crowd reports, satellite remote sensing Such, in particular, Information can be a useful tool for guidingtheplanningofUAVflightpathsoveragivenarea. inordertolocatethosewhoareinriskHowever,thereare difficultieswithcomputationaltractability.Whenplanning overtheresultingextremelylargeactionspaces,keepthis in mind. To address these issues, we define the survivor finding problem and offer the first example of a survivor discoveryalgorithmasoursolution.TheMonteCarlotree search method is a coordinated Monte Carlo tree search approach with continuous factors. Our assessment in comparison to Benchmarks reveal that our approach, Co CMCTS, is capable of localising more data. On simulations with real world data, casualties were 7 percent or more faster than typical procedures. Unmanned Aerial Vehicles (UAVs) that are low cost, reliable, and commercially available are becoming more common. (UAVs) has resulted in a concentrated attempt to use these platforms to assist first responders. obtaining sensory data without endangering human life.The concept of enabling coordinated UAVs to tour a catastrophe area in order to find the geographical location of victims is at the heart of this research. Given the huge region to cover and the continuous action space represented by a UAV's axes of motion,thisisa demanding assignment. Advancesindata collection have facilitated this study by opening up new sourcesofinformationondisasterscenarios,resultingina greater understanding of the situation on the ground during a disaster. Crowd sourced data is becoming more readily available due to the speed with which it may be created and its capacity to immediately represent the experiencesofthosepresent. People whocan often give a very accurate report on the hazards in their vicinity and thenumberofpotentialcasualties.

[19] The process of seeking for and rescuing missing individuals or other objectives in marine territory is knownasMaritimeSearchandRescue(MSAR).Thetypical workflow of a search and rescue (SAR) operation, including MSAR. A SAR operation begins when the rescue organisation is notified of a distress alert from any authority.Thelocationwillbeutilisedtodefinethesortof SAR operation that will be done, such as wilderness, urban, combat, or marine, and then information on the search area will be acquired based on location,

environment, and incident area size. The SAR team will choose the suitable search grid if the location of the subject(s) is unknown. Track line, parallel track, expandingsquare,sectorsearch,creepinglinesearch,and contour search are the six types of search patterns commonly utilised in marine SAR operations. As mentioned, the search here uses the parallel track search pattern(i.e.aSweepSearch/ParallelPatternconductedon a square/rectangular/cell grid), the search being conducted with a few Rotary Wing UAVs flying in formation. This can be visualised, indicating also the camera lineofsights of the UA Vs.Theflightpaths can be pre programmed.

4. CONCLUSIONS

The ability to consequently perceive people on drone photos utilising PC vision advances is a gigantic guide in SARtasks.Inthisexploration,weexaminedstateoftheart human recognition in drone photographs. On a few datasets, we explored the conduct of various CNN based article locators, including MobileNetV3 and YOLOv4 and soon.Asfarasnormalaccuracy,YOLOv4hasachievedthe bestdiscoveryresults(AP).InSARtasks,themodelought to have as not many bogus location (FPs) as practical to trynottosquanderassets.Inanycase,whilesearchingfor a missing individual, the most pivotal issue is that the locator finds that individual, and how exact the discovery islesshuge.JustgoforitoutflankedanyremainingAPsas farasitemsizeandlocationexactness.

REFERENCES

[1] A. Howard, M. Sandler, B. Chen, W. Wang, L. C. Chen, M.Tan,G.Chu,V.Vasudevan,Y.Zhu,R.Pang,H.Adam, and Q. Le, ‘‘Searching for 6 MobileNetV3,’’ in Proc. IEEE/CVFInt.Conf.Comput.Vis.(ICCV),Oct.2019,pp. 1314 1324.

[2] A.Al Kaff,M.Gómez Silva,F.Moreno,A.delaEscalera, and J. Armingol, ‘‘An appearance based tracking algorithm for aerial search and rescue purposes,’’ Sensors,vol.19,no.3,p.652,Feb.2019.

[3] K. Yun, L. Nguyen, T. Nguyen, D. Kim, S. Eldin, A. Huyen,andE.Chow,‘‘Smalltargetdetectionforsearch andrescueoperationsusingdistributeddeeplearning andsyntheticdatageneration,’’Proc.SPIE,vol.10995, May2019,Art.no.1099507.

[4] ”MultimodelDeepLearningforPersonDetectionin

[5] AerialImages”MirelaKundidVasi´candVladanPapi´c ,7September2020

[6] ”Unsupervised Human Detection with an Embedded Vision Systemona FullyAutonomousUAV for Search and Rescue Operations” Eleftherios Lygouras ,

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

Nicholas Santavas , Anastasios Taitzoglou , Konstantinos Tarchanidis , Athanasios Mitropoulos and Antonios Gasteratos, Department of Production and Management Engineering, Democritus University ofThrace,14August2019

[7] Detection of bodies in maritime rescue operations using Unmanned Aerial Vehicles with multispectral cameras Antonio Javier Gallego , Antonio Pertusa, Pablo Gil , and Robert B. Fisher ,Computer Science Research Institute, University of Alicante, San Vicente delRaspeig,03690,Spain

[8] A. S. Laliberte and A. Rango, ‘‘Texture and scale in object based analysis of subdecimeter resolution unmannedaerial vehicle(UAV)imagery,’’IEEE Trans. Geosci.RemoteSens.,vol.47,no.3,pp.761 770,Mar. 2009

[9] N.Tijtgat,W.VanRanst,B.Volckaert,T.Goedeme,and F.DeTurck,‘‘Embeddedreal timeobjectdetectionfor a UAV warning system,’’ in Proc. IEEE Int. Conf. Comput. Vis. Workshops (ICCVW), Oct. 2017, pp. 2110 2118

[10] O.Russakovsky,J.Deng,H.Su,J.Krause,S.Satheesh,S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein, A. C. Berg, and L. Fei Fei, ‘‘ImageNet large scale visual recognitionchallenge,’’Int.J.Comput.Vis.,vol.115,no. 3,Dec.2015,Art.no.211252

[11] M. Ivašić Kos, M. Krišto, and M. Pobar, ‘‘Human detectioninthermalimagingusingYOLO,’’inProc.5th Int. Conf. Comput. Technol. Appl. (ICCTA), New York, NY,USA,Apr.2019

[12] M. Ivasic Kos, M. Kristo, and M. Pobar, ‘‘Person detection in thermal videos using YOLO,’’ in Proc. SAI Intell.Syst.Conf.Cham,Switzerland:Springer,2019

[13] D.Du,Y. Qi, H.Yu,Y. Yang, K.Duan,G. Li,and Q.Tian, ‘‘The unmanned aerial vehicle benchmark: Object detection and tracking,’’ in Proc. Eur. Conf. Comput. Vis.(ECCV),2018

[14] M. R. Hsieh, Y. L. Lin, and W. H. Hsu, ‘‘Drone based object counting by spatially regularised regional proposal network,’’ in Proc. IEEE Int. Conf. Comput. Vis.,Oct.2017

[15] M. Mueller, N. Smith, and B. Ghanem, ‘‘A benchmark and simulator for UAV tracking,’’ in Proc. Eur. Conf. Comput.Vis.Cham,Switzerland:Springer,2016

[16] Very Deep Convolutional Networks For Large Scale Image Recognition, Karen Simonyan & Andrew Zisserman,2014

Volume: 09 Issue: 06 | Jun 2022 www.irjet.net p ISSN:2395 0072 © 2022, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal |

[17] Deep Residual Learning for Image Recognition,Kaiming He Xiangyu Zhang Shaoqing Ren JianSun,June 2016

[18] Feature Pyramid Networks for Object Detection Tsung Yi Lin , Piotr Dollar´ , Ross Girshick , Kaiming He,BharathHariharan,andSergeBelongie,Jul.2017

[19] “PlanningSearchandRescueMissionsforUAVTeams” Chris A. B. Baker and Sarvapali Ramchurn and W. T. LukeTeacyandNicholasR.Jennings

[20] Determining Position of Target Subjects in Maritime Search and Rescue (MSAR) Operations Using Rotary Wing Unmanned Aerial Vehicles (UAVs) ,Zaharin Yusoff Faculty of Defence Science and Technology National Defence University of Malaysia Sungai Besi Camp57000KualaLumpur,Malaysia