ACCIDENT DETECTION USING BiLSTM
Abstract - Accidents have consistently ranked as the major cause of death in India. More than eighty percent of the fatalities that occur as a result of accidents are not directly attributable to the accident itself; rather, they are the result of victims not receiving prompt assistance. It is possible for an accident victim to be left unattended for a significant amount of time on routes that have very light and quick traffic. The objective is to design a system that is able to determine whether or not an accident has occurred based on the video input received by the system. It is the intention to run each frame of a video through a convolutional neural network and BILSTM models that have been trained to identify video frames as either accident or non-accident frames. The Convolutional Neural Network and the BiLSTM models have been shown to be a method that is both quick and accurate when it comes to identifying photographs. CNNbased image classifiers have attained an accuracy of greater than 95% with fewer datasets, and they require less preprocessing than other image classification techniques.
Key Words: Convolutional neural network and BILSTM
1.INTRODUCTION
The main goal is to implement a system that can recognize an accident from provided video material. The systemismeanttobeatooltoassistanaccidentbypromptly identifying an accident and afterward reporting the authoritiesaboutit,youcanhelpthoseinneed.Thegoalisto usecuttingedgeDeepLearningAlgorithmsthatuseBILSTM andConvolutionalNeuralNetworks(CNNsorConvNet)to analyzeframescapturedfrom thevideoinputgiventothe systeminordertoidentifyanaccidentwithinsecondsofit occurring.Weconcentratedoninstallingthistechnologyon highways where there is less congestion and prompt assistanceforaccidentvictimsisuncommon.Transportation isalegitimatemeansoftakingorcarryingitemsfromone locationtoanother.Astimegoeson,transportationsuffersa numberofproblems,includingahighaccidentrate,traffic jams,airpollutionfromcarbonemissions,andmore.
The transportation industry occasionally struggled with reducingtheseverityofcrash-relatedinjuriesinaccidents. Since transportation is so complex, researchers have
***
combined virtual technologies with it to create the IntelligentTransportSystem.Inthesphereoftransportation, theconceptofintegratingvirtualtechnologiesisinnovative, and it is essential for resolving problems in a worldwide context.Thetraditionalmethodforcreatingnext-generation technology is known as ITS. From ITS implementations, a varietyofreimbursementsareavailable.ITScansignificantly lowerhazards,accidentrates,trafficjams,carbonemissions, and air pollution and meanwhile, improving all modes of transportation's traffic flow, transit speeds, and levels of passengersatisfaction.OneofITS'skeyusesistrafficcontrol. Controlling traffic is becoming a major challenge as the overcrowding issue gets worse. The Video traffic surveillancesystemisoneofthekeytechnologiesbeingused toimplementsolutionsforthisproblem.
2. RELATED WORK
[1] Asanalyticaltoolsinthisparticularstudyproject,CNN, RNN, and LSTM were employed. Four layers make up the study'sarchitecture:twoconvolutionallayersthathelpwith feature extraction, two layers of long short-term memory (LSTM)units,andatoplayer.LSTMisinchargeofcontrolling eachvideo'stimedependence(LongShort-TermMemory). Over80%accuracyinvalidationisachieved,withthesheer amountofdatabeingoneofthemainchallenges.
[2] TransferLearningandMaskR-CNN,whichusesamaskRCNNtodetectcars,arethemaintechniquesthatweemploy inthisstudy.Theintersectionoverunion(IoU)algorithmis used in order to discover collisions. If used in conjunction withastrongResponsesystem,thismodelcouldreducewait times,speedupprocedures,andimprovedetectionaccuracy.
[3] The cornerstone of the accident detection system is providedbytheCVISandmachinevision.Wedesignedthe YOLO-CAdeepneuralnetworkmodeltodiscoveraccidents. DeeplearningtechniquesandCAD-CVISwereusedtocreate thismodel.Weusealossfunctionwithdynamicweightsand Multi-Scale Feature Fusion (MSFF) to enhance the recognitionaccuracyofverysmallobjects.Whenitcomesto choosing proposal regions, Fast R-CNN uses the timeconsumingselectiveresearchapproach.Whendealingwith very large objects, rapid R-detection CNNs give incorrect positivefindings.
[4] Theresultsofthisstudysuggestthatthebestapproach to the issue is to apply video analytics techniques. The structureiscomposedoftwodistinctcomponents.Thefirst
one uses a modified version of the architecture from Inception V4 to extracta vectorof visual attributes. Then, we'll discuss the following two steps: temporal video segmentationandautonomoustrafficaccidentrecognition. Thefeaturevectorproducedbyutilizingamodifiedversion oftheInceptionV4architectureisacceptedasinputbythe authors' proposed neural network architecture, which is constructedonaConvLSTMlayer.Duetothelackofeasily accessible cases involving pedestrians, cyclists, and motorcyclists, the technique can only be used in crashes involving motor vehicles with higher performance in the recognitionoftrafficcrashescapturedbyvideo.
[5] ThecapabilitiesofaGPSreceiverwillbeusedwithinthe parameters of this inquiry. The phrase GPRMC will be detectedonceeverysecondbyaGPSdevice.Tocomparethe new and old velocities, the MCU was employed. If the computerisoutfittedwithaGSM/GPRSmodem,itisfeasible forthecomputertoreadthedataandtextmessagesthatare transmitted by GPRS. used a GSM modem, which is quite readilyavailableandwell-liked.Sendsthecar'smostrecent recordedspeed, whichcanbeusedtodeterminehowbad thecollisionwasandifitwasprogrammedto,startanaudio call.
[6] ThisprojectrequirestheuseofaRaspberryPi3Model B+, a GSM Module SIM800L, and a Pi Camera. The recommended car accident detection system has the capacity to track accidents as they occur in real-time and immediately text relevant medical facilities and law enforcement agencies to inform them of the incident. The suggested alternative is also more cost-effective than the currentmethods,whicharemoreexpensiveandlessreliable since they rely on expensive sensors and unnecessary technology.
[7] In this article, a system for identifying instances of vehicle collisions in egocentric films using unsupervised deeplearningisproposed.Themethodologiesusedbythis system are trajectory prediction, which uses LSTM for pedestrian trajectories and their interactions, and video anomaly detection, which primarily targets video surveillance scenarios and typically uses an unsupervised learningmethodforthereconstructionofregulartraining data. On the reconstruction of typical training data, these two strategies are referred to as unsupervised learning methods.Itproducespredictionsabouttheroutestakenby traffic participants and their future positions, and it leveragestheconsistencyandaccuracyofthosepredictions asproofthatanunforeseeneventmighthaveoccurred.The accuracyandconsistencyofthisstrategy,whichpredictsthe trajectories of persons taking part in the traffic as well as their future locations, are employed as indicators that an aberrationmayhavehappened.
[8] In this particular article, Automatic Smart Accident Detection(ASAD)technologyisused.AccidentDetectionand AlertingDevice(ASAD)isaservicethatmaybeinstalledin
automobiles and is activated in the case of an accident. Mamdani fuzzy logic can be used to detect accidents. axis accelerometer and gyro breakout for the MPU-6050. This element is used to gauge the vehicle's rotation and acceleration.ThearchitecturealsohasfourForceSensitive Resistors (FSR) that measure the force of an accident's impact and are connected to the vehicle's four ends.This system provides a service that automatically alerts local authoritiestoanyeventsthathaveoccurredintheircities. The outcome is that the authorities can respond to the problemrightaway.Asmuchasyoucan,preventharmfrom comingtothepopulaceandtheeconomy.
[9] Using VANET (Vehicular Ad-hoc Network), vibration sensors, and piezoelectric sensors, traditional accident detection techniques are used in this system. b) Machine learningandartificialintelligence-basedaccidentdetection methods, such as support vector machine accident prediction and fuzzy logic accident detection. c)Hybrid methodsusinglimitswitches,mobilephonewearinessand intoxication detection, and accelerometer speed and acceleration measurement. For the goal of accident detection, the system in question made use of numerous sensors, such as accelerometer sensors, shock sensors, pressuresensors,etc.,aswellasnumerousmachinelearning techniques, such as neural networks, support vector machines,representationlearning,etc.
[10] Inthiscase,theprototypewasbuiltandthenputintoa remote-controlled toy car. This system uses GSM and GPS technologies,aswellasvehicleadhocnetworksandmobile applicationsensors.Italsohasaheartratemonitor.Assoon asanaccidentisdetected,theheartratesensorfindsoutthe driver'sheartrateandtheGPSmodulefindsoutwherethe driver is. An SMS is then sent to the driver's emergency contacts.Whenthevehicleisinanaccidentthattipsitover ortiltsitmorethan30degreesandtheresetbuttonisnot pressed within the time limit, the system will send the messagetotheemergencynumbersthathavealreadybeen saved.
[11] The suggested approach uses machine learning algorithmsinstalledineachcartoworktogetherwithother vehiclesthatareoutfittedwithV2Vcommunicationdevices to predict the likelihood of accidents. The underlying workingsofthreedifferentmachinelearningtechniques artificialneuralnetworks(ANNs),supportvectormachines (SVMs), and random forests are examined in this article (RFs).SUMO(SimulationofUrbanMobility),acollision-free traffic controller, is currently being used on roads in an efforttoloweraccidents.
[12] UsingtheMATLABandSIMULINKsoftwarepackages, this model is developed that has been presented is constructed and tested. The following are the primary components are Detection system using DWDC(Dynamic Webester Dynamic Cycle) to reduce the amount of time spent waiting and to improve the flow of traffic accident
detectionTheplannedhybridtransportationsystemoffered ananswertotheproblemoftrafficcongestion.Thismodel cut down on the amount of time spent waiting at traffic signals, which resulted in time savings for drivers. As the subsystems collaborate and share data with one another, there is a possibility that some of the flows will contain redundantorunnecessarydata.
[13] Thespecifiedapparatusmakesuseofacomplementary set of HCSR04 ultrasonic sensor modules. Within this automobile, there are sensor modules located in both the front and the rear windscreens. Both of these units are mountedinthesidewindowsoftheautomobile.Afterthen,a calculationismadetodeterminethedistancebetweenthe sensorunitsandthebumpers.The"firstthresholddistance" and the "second threshold distance," respectively, are the termsthatareusedtorefertothesedistances.
When moving away from the car, everything is alwaysatagreaterdistancethanthethresholdsthatthecar has set. This occurs whenever something is moving away from the path that the car is travelling on. If something strikesthevehiclehardenough,itwilltravelfurtherthanthe predetermined threshold distance, which will activate the processing system. The technology quickly calculates the location of the car using GPS and then transmits that information to the relevant authorities through GSM. The systemthathasbeenproposedwilloperateinthismanner. However,theprocedurethathasbeendescribedissufficient fordeterminingwhetherornottherehasbeenacollisionon theroad.However,thereareanumberofproblemswiththe remedythathasbeenoffered.TheHCSR04soundsensorhas amaximumdetectiondistanceoffourmetersinanygiven direction.Therefore,vehicleswithathresholddistancethat islargerthanfourmetersareunabletoutilizetheproposed technique. Since the ultrasonic sensor module can only detect reflected sound waves within a range of fifteen degrees, the location of the sensor module plays an importantroleindeterminingthequalityofthediscoveries that are produced by the module. It's possible that an incorrectlypositionedsensoristofaultforthefalsealarmin thisinstance.
[14] Locatingmovingcarsisthefirststepinthesuggested method.Thisisdonebyfirstextractingtheforegroundwith GMM(GaussianMixtureModel),andthengoingontomotion mapping. The intensity of the car's motion as well as the direction in which the car is moving are then used to establish whether or not a collision has taken place. Find eachandeveryvehiclethatisparkedintheparkinglot.The accuracyofthemethodforidentifyingaccidentscenarioshas been significantly improved by the utilization of the AND operatorforthepurposeofmerginginformationfromthe foreground and the motion map. Greater than 75% of collisionsinvolvingautomobilesmaybecorrectlyidentified usingthisapproach.
[15] Vehicledetection,tracking,andparameterextraction are the three distinct tasks that make up the proposed technique, in that order. The three primary systems that worktogethertodetectcarsaretheGaussianMixtureModel (GMM), Mean Shift Algorithm, and Accident Detector. The meanshiftapproachisusedtofollowtheobservedcarsafter theyhavebeenidentified.Thismethodhandlesocclusions duringaccidentsreasonablywell,butithasasevereproblem in that it depends on a small set of parameters, making it difficulttoadapttosituationslikesuddenchangesintraffic patternsorinclementweather.Thisparadigmisbased on local parameters such as trajectory intersection, velocity calculation, and the anomalies related to these. The recommendedframeworkiscapableofcorrectlydetecting accidents, as evidenced by the 71% Detection Rate and 0.53%FalseAlarmRateonaccidentrecordingsacquiredin variedsettings.However,duetofaultsinvehiclerecognition and tracking, this method is not suitable for high-density traffic. Be at ease; these errors will be corrected in subsequentwork.
Largeobjectsinthecameras'fieldofviewmayalsohavean impact on how well they follow the vehicles, which may thereforehaveanimpacton howwelltheydetectcrashes snow,day,night,andmanyweathersituations.
[16] Long short-term memory (LSTM) and convolutional neuralnetwork(CNN)layersareusedinthismodeltolabel real-timevideomaterial.ThefootagecapturedbytheCCTV cameras is immediately transmitted to a component responsibleforpre-processing.
Thefirststageofprocessing,whichinvolvesextractinghighquality still images from video, is the responsibility of the openCV library. Some changes will be made to the dimensionsandshapesofthephotosinorderforthemtobe compatiblewiththeResNet-CNNalgorithm.Whencompared toothersystems,thismethodisfarsuperiorintermsofcost, durability, and accuracy. Making changes to a piece of software in time for it to be implemented in a real-world scenariocanbeadifficulttask.
[17] The system proposed is used "to detect using video" and "to detect using audio" are the two models that are producedasaresultofthis.WhileSVMandCNNwereboth utilizedthroughouttheprocessoftestingtheclassificationof thevideoinputs,onlyCNNwasutilizedduringtheprocessof evaluatingtheclassificationoftheaudioinputs.ManyGRU layers are utilized throughout the process of the training method. CNN filters are applied to each of the three dimensions of the data before it is classified. In this approach,multiplemodesofinquiryarecombinedinorder to get information that is more specific. Models that can handleonlyasinglecategoryofdatashouldbeavoidedin favourofthesestrategies.Themostsignificantdrawback,on theotherhand,isthatproducingCNNcontentin3Dtakesa greatdealmoretimethanitusedto.
[18] Residual Networks, also referred to as "ResNets" for short,areacommontypeofneuralnetworkthatformthe basisfornumerouscomputervisionapplications.Extraction of key frames, extraction of features, clustering, and classificationarethe four primarypillarsthatsupportthe system. And it does it in a timely and accurate manner, identifyingwheretheerroroccurred.Thissystemismore efficientintermsofcostwhencomparedtoothermethods.
[19] Amotorcycleaccidentdetectionandalertsystemthat takes into consideration the vehicle's acceleration, deceleration,tilt,andchangesinthepressurethatisbeing appliedtothevehiclebody.Combiningaglobalpositioning system (GPS) with a proximity sensor results in an intelligent distributed system that has the potential to identifyaccidentsandalerttheappropriateauthorities. In addition to this, it is unable to recognize two-wheeler incidentsaseffectivelyasothersystemscan.
[20] Anaccelerometerallowsa caralarmtodetect errant drivingbehaviorsolongasthedriverispayingattention.It can serve as a crash or rollover detector in the case of an accidentandbeusedinthiscapacityifnecessary.Thesignal ispickedupbyanaccelerometer,whichthencalculateshow severethecollisionwasbasedonitsreadings.
Iftherewerenoinjuriesandthereisnoimmediatethreat, thedrivermayuseaswitchtoturnoffthealarmmessageif itisnolongernecessary.Inadditiontoauserinterfacethat is easier to navigate and a foundation of reliability that is morerobust,thelevelofsensitivityandaccuracythatcanbe achievedhassignificantlyimproved.
3. CONCLUSIONS
Accidents are one of the most prevalent sorts of challengesthathumanityfacesonadailybasis,andtheycan leadtothelossoflifeaswellasthedestructionofmaterial goods. The strategy that has been suggested provides a solution to this problem that is one that is realizable and effectiveatthesametime.Thesystemfortheidentification ofautomotiveaccidentsthathasbeendevelopedisableto monitor the situation from the moment an accident takes place until it is resolved. Comparatively speaking, the proposedsystemismuchmoreaccurate,cost-effective,and reliable than its competitors. This is primarily because a model-based approach is used instead of the expensive sensors and unnecessary hardware used in other systems thatarealreadyinuse.Theproposedsystemalsohasamuch higher accuracy rate. A higher level of sensitivity and accuracy is indeed feasible utilizing this technology, accordingtoexperiments,tests,andvalidationsthathaveall been carried out with the use of images. In each of these procedures,imageshavebeenutilized.Asaconsequenceof this, deploying this system over the bulk of the country's stateandnationalroadwaysisapossibilitythatshouldnot bediscounted.Throughoutthephasesofexperimentation, testing,andvalidation,imageshavebeenutilized.
REFERENCES
[1] Durgesh Kumar Yadav, Renu, Ankita, Iftisham Anjum, "Accident Detection Using Deep Learning",2020, 2nd International Conference on Advances in Computing, CommunicationControlandNetworking(ICACCCN).
[2] Akshit Diwan, Vandit Gupta and Chaitanya Chadha, “AccidentDetectionUsingMaskR-CNN”,International JournalforModernTrendsinScienceandTechnology, Vol.07,Issue01,January2021,pp.-69-72
[3] Daxin Tian, (Senior Member, IEEE), Chuang Zhang, Xuting duan and Xixian wang, "An Automatic Car Accident Detection Method Based on Cooperative VehicleInfrastructureSystems",2019,10.1109/ACCESS.2 019.2939532.IEEE Access. K. Elissa, “Title of paper if known,”unpublished.
[4] SergioRobles-Serrano,GermanSanchez-TorresandJohn Branch-Bedoya,"Automatic Detection of Traffic Accidents from Video Using Deep Learning Techniques",2021,https://www.mdpi.com/2073431X/1 0/11/148
[5] Md. Syedul Amin, Jubayer Jalil,M.B.I.Reaz,"Accident Detection and Reporting System using GPS,GPRS and GSMTechnology",2012,IEEE.
[6] .Sreyan Ghosh,Sherwin Joseph Sunny,Rohan Roney,"AccidentDetectionusingConvolutionalNeural Networks",2019,IEEE.
[7] YuYao,MingzeXu,YuchenWang,DavidJ.Cranda,EllaM. Atkins,"UnsupervisedTrafficAccidentDetectioninFirstPersonVideos",2019,IEEE.
[8] AsadAliandMohamadEid,"AnAutomatedSystemfor AccidentDetection",2020,IEEE.
[9] Unaizaalvi,MuazzamA.KhanKahattak,(SeniorMember, IEEE),Balawal Shabir, Asa Waqar Malik, (Senior Member, IEEE),and Sher Ramzan muhammad, "A ComprehensiveStudyonIoTBasedAccidentDetection Systems for Smart Vehicles",2020,Digital Object Identifier10.1109/ACCESS.2020.3006887.
[10] Nagarjuna R Vatti,PrasannaLakshmi Vatti,Rambabu Vatti,Chandrashekhar Garde"Smart Road Accident DetectionandcommunicationSystem",2022,IEEE.
[11] Nejdet Dogru,Abdulhamit Subasi,"Traffic Accident DetectionUsingRandomForestClassifier",2019,IEEE.
[12] Abdulrahman Abdullah Alkandari,Meshari Aljandal,"TheoryofDynamicFuzzyLogicTrafficLight Integrated System with Accident Detection and Action"2015,IEEE.
[13] Usman Khalil,Adnan Nasir,S.M. Khan,T.Javid,S.A.Raza,A.Siddiqui"Automatic Road AccidentDetectionusingUltrasonicSensor",2019,IEEE.
[14] DoKyungShin,WooJinJeong,JongMinLee,KiTaePark andYoungShikMoon,"Caraccidentdetectioninparking lots"2019,IEEE Transactions on Intelligent TransportationSystems,vol.1,no.2,pp.108-118,2022.
[15] YuChen,YuanlongYu andTingLi,"AVisionbasedTraffic Accident Detection Method Using Extreme Learning Machine",2014,IEEE.
[16] Abishek T S,2Adith Jose,3Albin Davis,Alan Thomas,JasmyDavies,"AccidentDetectionandWarning System",2021,IEEE.
[17] Jae Gyeong Choi,Chan Woo Kong,Gyeongho Kim,SunghoonLim,"Carcrashdetectionusingensemble deep learning and multimodal data from dashboard cameras",2020,www.elsevier.com/locate/eswa
[18] Aman Kumar Agrawal, Kadamb Agarwal, Jitendra Choudhary, Aradhita Bhattacharya, Srihitha Tangudu, Nishkarsh Makhija and Rajitha B,"Automatic Traffic Accident Detection System Using ResNet and SVM",2020,IEEE.
[19] Amit Rathode,Sujeet Kumar,Sunny Shankar,Chinmay Tapkire,Sonam Borhade,"Accident Detection and ReportingUsingDeepLearning",2021,IEEE.
[20] S.Parameswaran, P.Anusuya,M.Dhivya,A.Harshiya Banu,D.Naveen Kumar,"Automatic Vehicle Accident DetectionandMessageingSystem",2018,IEEE.