基于数据关联的多目标跟踪算法研究.pdf

ABSTRACT

Multi-objecttracking,asanimportantresearchsubfieldofcomputervision,aimsto

maintaintheidentityandtrajectoryofeachtargetinacontinuousimagesequence,soit

iswidelyusedinreallife,suchasunmanneddriving,intelligentsecurity,industrial

intelligence,etc.Nowadays,multi-objecttrackingismostlytrackbydetection,which

mainlyreliesontheobjectdetectortoprovidethepositioninformationoftheinterested

targetinthecurrentframe,andmatchesthetrackingtargetandthedetectiontarget

throughdataassociation,soastocompletethemulti-objecttracking.

Thisthesisstudiesthemotionmoduleandappearancemoduleindataassociation-

basedmulti-objecttracking.ByproposingaUnitedSiameseNetwork,themotionmodule

andappearancemoduleinthedataassociationareregardedasawholetoestablishthe

framework.Then,theobjectdetectionmodule,motionmoduleandappearancemodule

arefurtherintegratedintoanend-to-endnetwork,andamulti-objecttrackingalgorithm

whichintegratingdetectionandtrackingisconstructed.Atthesametime,fortheairport

scenesurveillance,anintegrateddetectionandtrackingalgorithmbasedonADS-Bis

designedbycombiningthisalgorithmwiththeADS-Bsignalontheaircraft.Finally,the

experimentsshowthattheseworkshaveachievedgreatimprovementcomparedwith

otheralgorithms.Thespecificcontentandinnovationoftheworkareasfollows:

1.Thisthesisfirstintroducesthebasicsofdeeplearningandseveralclassicdeep

learningnetworksandexcellentobjectdetectionalgorithmsinrecentyears.Atthesame

time,someclassicmethodsofmotionmodule,appearancemoduleandassociation

algorithminmulti-objecttrackingalgorithminrecentyearsaresummarizedandthe

problemsexistingbetweenthesemodules