基于改进的ResNet网络和特征融合的自标跟踪算法

打开文本图片集

Object tracking algorithm based on improved ResNet network and feature fusion

MENG Weijun'，SUN Siwei'，MA Sugang1，² （1.SchoolofomputerScienceandTechnologyXi’anUniversityofPostsandTelecommunications，Xi'an7，Cina; 2SanxiKeytselii

Abstract：Anobject trackingalgorithmbasedonimprovedResNetnetworkand feature fusion hasbeenproposedon the basisoftheATOM5Oalgorithm，soastoenhancetheobjectfeaturesextractedwithresidualnetworks.Anenhancedbotteneck block，integratingbatch-freenomalzation（BF）andposition-awareircularconvolution（ParC），ismploedwitintheReset-50 backbonenetwork.Thisefectivelybolstersthecaptureofglobalinformationandmitigatestheaccumulationoftrackingdrift.For theextractedfeatures，theatentionfeaturefusionmoduleisusedtofurtherenhancetheexpressionabilityofthefeaturesforthe objectbyfusingthedetailsofshallwfeaturesandthesemanticinformationofdepfeatures.Theproposedalgorithmwas validated with the OTB2015，VOT2018 and LaSOT datasets.Itachieves a success rate and precision of 70.2% and 91.1%， respectively，ontheOTB2015dataset，whichisimprovedby1.2%and1.5%，respectively，overthebenchmarkATOM50 algorithm. On the VOT2O18 dataset， its expected average overlap rate saw an increase of 4.4% . On the LaSOT dataset，its success rateand precisionareimprovedby2.4%and2.9%，respectively.Itsaverage tracking spedontheOTB2015dataset reaches 34.3 f/s，ensuring real-time tracking.

Keywords：deep learning;visual tracking; Siamese network;batch normalization;attention mechanism;improved ResNet network

0 引言

目标跟踪是计算机视觉领域的核心问题之一，其目的是给定第一帧的单个目标，在后续帧中找到与被跟踪目标最匹配的区域，并以矩形框的形式表示目标的大小和位置。（剩余14508字）