An Efficient Direction Field-Based Method for the

Sensors2011,11,7364-7381;doi:10.3390/s110807364sensorsISSN1424-8220*SchoolofElectronic,InformationandElectricalEngineering,ShanghaiJiaoTongUniversity,Shanghai200240,China;E-Mails:[e-mail](J.Y.);[e-mail](W.T.);[e-mail](M.L.);[e-mail](Y.Z.);[e-mail](H.Z.)*Authortowhomcorrespondenceshouldbeaddressed;E-Mail:[e-mail];Tel.:[phone];Fax:[phone].Received:28June2011;inrevisedform:14July2011/Accepted:15July2011/Published:25July2011Abstract:Railwayinspectionisanimportanttaskinrailwaymaintenancetoensuresafety.Thefastenerisamajorpartoftherailwaywhichfastensthetrackstotheground.Thecurrentarticlepresentsanefficientmethodtodetectfastenersonthebasisofimageprocessingandpatternrecognitiontechniques,whichcanbeusedtodetecttheabsenceoffastenersonthecorrespondingtrackinhigh-speed(upto400km/h).TheDirectionFieldisextractedasthefeaturedescriptorforrecognition.Inaddition,theappropriateweightcoefficientmatrixispresentedforrobustandrapidmatchinginacomplexenvironment.Experimentalresultsarepresentedtoshowthattheproposedmethodiscomputationefficientandrobustforthedetectionoffastenersinacomplexenvironment.Throughthepracticaldevicefixedonthetrackinspectiontrain,enoughfastenersamplesareobtained,andthefeasibilityofthemethodisverifiedat400km/h.Keywords:machinevision;fastenerrecognition;DirectionField(DF);LDA;patternmatching;VOSSLOHfastener1.IntroductionNowadays,largerailwaynetworksneedstrongsecuritymechanismsandmustbemaintainedcontinuouslytoensuresafety.However,forbothheavyworkandobjectiverequirements,trainedpersonnelforrailwaymaintenancecannotmeettheneedsofperiodicinspections.Therefore,severalOPENACCESSSensors2011,117365automaticraildetectiondevices[1-3]areused,suchasJapan’sEast-I,Austria’sEM-250,Germany’sRAILAB,Italy’sRoger2000,France’sMGV,theAmericanEnsco,andImageMap’strackinspectioncar,whichareallknownforrailwaymaintenance.Anumberoftasksareinvolvedinrailwaymaintenance,suchasgaugemeasurement[4],trackprofileandwearmeasurement[5,6],ballastmeasurement[7],raildefectdiagnosis[8],andconcreterailseatabrasion,amongothers.Thecurrentarticlefocusesonfastenerdetectionasanimportantpartofrailinspectionsystems[5],whichsecurestherailtothesleepersorconcretebase.Fastenerdetectionisanimportantserviceinvolvedinrailinspection.Ouraimistodetectwhetherfastenersarepresentattherelevantlocationinahigh-speedrailway.Inpracticalsituations,themainkindoffastenerusedistheGermanVOSSLOH,asshowninFigure1(dashedline).Figure1.Sampleimageoffasteners.Thecurrentliteratureprovidesafewimage-basedmethodsforfastenerdetection.Hsiehetal.[9]appliedamorphologicalapproachtoobtaintheoutlineoffasteners,whichcaninspectthemontheconcreteorballastedtracktodetermineiftheyarebrokenandwhichhasa77%recognitionrateforbrokenclips.Itentailsmuchcomputationtimebecauseofitstime-consumingimageprocessingtaskssuchashistogramanalysis,openandcloseoperation,etc.However,itcannotsuppresslargebrightnessvariancesandenvironmentinterferencesandhencecannotmeetreal-timerequirements.Stellaetal.[10]andMarinoetal.[11]usedwaveletanalysisandmultilayerperceptronneuralclassifierstodetecthexagonal-headedbolts(akindoffastener),revealingthepresence/absenceoffastenerboltswithhighdetectionandclassificationrates.However,theresultisnotapplicabletooursituationduetothecomplexshapeandenvironmentoftheVOSSLOHfastener.ForVOSSLOHfastenerspecificdetection,thecurrentstudyaimstocompleteafastenerdetectionsystemwithcomplexshapesinacomplexenvironment.Itintroducesarapidfastenerdetectionmethodthatcompleteshigh-speednoncontactdetectionbasedonthemachinevisiontechnique.Theenvironmentofthefasteneriscomplex,soitsimagesareunstable.Sometimes,theedgeofelasticclipsforthefastenerisobscured(asshowninFigure1)becauseofsurfaceerosion,dustcoverage,motionblurring,andbrightnessfluctuations.Further,stonesandotherdebristhatcauseocclusionsarefound.Therefore,simplebrightnessandedgefeaturesarenotappropriateforrecognition.Regionalfeaturesaredifficulttoobtain.Afterextensivetesting,wefoundthatacertainscaletextureofthefastenerisstable,whichisindependentofbrightnessandrobustforpartialSensors2011,117366occlusion.Hence,weintroduceaDirectionField(DF)-basedmethod,whichiscommonlyusedtodescribetexture.ThecurrentpaperusesaDFtemplateofthefastenersforrecognition,introducingLinearDiscriminantAnalysis(LDA)toobtaintheweightcoefficientmatrixformatching.Tolimittheneedforcomputationwhileguaranteeingstability,weuseblockDFwithdiscretesamplingpoints.Thepresentworkisorganizedasfollows.Section2presentsthemainsystemoverview.Section3focusesonsomeDFtheoriesrelevantforourapplication.Section4discussesthematchingmethod,andSection5presentsanumberofteststoverifythefeasibilityoftheDFdescriptorandthematchingstability.Section6introducesthedisturbanceexperimentandthepracticalrailwayimageresults,andfinally,Section7presentstheconclusionsandfut