Automatização do reconhecimento de buracos em rodovias usando inteligência computacional

Abstract

Potholes inasphaltpavementsandhighwaysareawell-knownproblem,whichisag- gravatedbythegrowingdevelopmentofautonomousvehicletechnology.Thistechnology, therefore, needstoincorporateefficientautomaticpotholedetectionsystems.Infact,there are differentmethodsforautomaticpotholedetection,however,thesemethodsstillneedtobe further evaluatedbothintermsoftheiraccuracyandspeedofresponseinordertobeusedin real contexts.Inordertocontributewithastepinthedirectionofthisevaluation,thiswork presents theproposalofasolutionforthedetectionofpotholesinimages,beingthedetection of theexistenceofapotholeandthedeterminationofthesameintheimage,theobjectiveof this work.Theworkhastwomainparts,thefirstbeinganimageclassificationarchitecture, which usestheHistogramofOrientedGradients(HOG)techniqueasanextractionofimage characteristics, andtwoclassifiers;ArtificialNeuralNetworksandSupportVectorMachines. The objectiveistooptimizeparametersanddeterminethebestclassifierfortheproblem.The second partpresentsaproposalforarchitecturetodetectpotholesinimages,determiningnot only theexistencebutalsothelocationofapotholeintheimage.Thistaskusestexturede- scriptors from Haralick in asystemof grid applied totheimages,addedtotheclassification architecture ofthefirstpart.Theprocessingtimesofthesearchitectureswerealsoevaluated. The resultspresentedshowthatarchitectureusingtheArtificialNeuralNetworkclassifieristhe best optionforpotholedetection,reaching83% IoU (IntersectionofUnion) in additionto73% accuracyand93%precision.ProcessingtimeisalsofavorabletotheNeuralNetworkclassifier reaching about9 frames per second(FPS).