Nombre: HEITOR DELESPORTE CONCEGLIERI

Fecha de publicación: 25/07/2023
Supervisor:

Nombre	Papel
ANSELMO FRIZERA NETO	Advisor *
RICARDO CARMINATI DE MELLO	Co-advisor *

Junta de examinadores:

Nombre	Papel
ANSELMO FRIZERA NETO	Advisor *
CAMILO ARTURO RODRIGUEZ DIAZ	Internal Examiner *
FAUSTO ORSI MEDOLA	External Examiner *
RICARDO CARMINATI DE MELLO	Co advisor *

Sumario: Thenumberofvisuallyimpairedpeopleintoday’ssocietyhasincreasedandvisionisof paramountimportancefortheindividual’sperception,orientationandlocationinthe environment.Devices,suchastraditionalcanesandguidedogs,andsmartcanes,which useultrasonicandinfraredsensors,areimplementedtoprovidesupporttousers.However, thesedeviceslackthemeanstoacquireimportantinformationtohelpinsituations commonlyreportedbyusers,suchasthedetectionandpositioningofhighobstacles.The presentstudyaimstopresentandvalidateanobjectdetectionstrategythroughimages capturedbyanRGB-Dcamera,dividedintotwostages.Inaddition,challengesand potentialitiesarealsodiscussedinthesynergisticimplementationofsuchastrategyto thedevicesalreadyusedtoaidlocomotion.TheuseofRGB-Dcamerasthenprovides relevantinformationthroughtheirdepthandRGBsensorstocomplementtheuseof traditionaldevicesandhelpusersmoveandnavigate.Asvariablesystemparameters (e.g.,theazimuthandpolarangles,andthedepthparameter)canconsiderablyaffectthe systemoperation,twogroupsofexperimentsarecarriedouttovalidatethestrategyand understandtheimpactsonthequalityofdetectionandoperationofthesystemduring navigation.Initially,thesystemisimplementedthroughthefirststageofthestrategy, withtheazimuthpresentingavariationratefrom0°to0.21ºandthedepthparameter, 1mmto14mm.Forthesecondgroupofexperiments,thesecondstepofthestrategy isemployedandanewdatabaseisgeneratedwithanewtrainedclass.Thetrainingof thenewclassallowsthesystemtodetect100%oftheimagesofthetrainedclasswitha minimumconfidencethresholdof0.887,indicatinggooddetectioncapability.Thesecond stageofthestrategyisabletodetectstaticanddynamicobjects,aswellaselevatedobjects andonthesameaxisasthecamera.Resultsoftheseexperiments,suchasthestandard deviationsofthepolaranglerangingfrom0.3°to0.34°,showtheviabilityoftheproposed detectionstrategyforbothstaticanddynamicscenarios,offeringalternativesforfurther studiesinrelationtoobjectdetectionformobilityassistanceforthevisuallyimpaired. Thegenerationofnewdatasetsfortheimplementationofthesysteminnewscenarios, suchastheurbanone,theintegrationindifferentmobilecomputationalplatformsaiming atpracticalityandmobility,andtheimplementationofotherfeedbackmethodstothe useraresomeofthefutureworksresultingfromthisresearch. Keywords:ComputerVision,Real-timedetection,Visuallyimpaired,RGB-Dcamera, Assistivetechnology.

Acceso al documento