基于深度光流估计网络的流速测量方法

打开文本图片集

关键词：深度光流估计网络；流速测量；循环全对场变换；特征增强；真实图像对渲染；速度转换模型中图分类号：TN911.73-34；TP391 文献标识码：A 文章编号：1004-373X（2026）01-0095-07

Flowvelocitymeasurementmethodbasedondeepopticalflow estimationnetwork

MA Yuekun1，2，3，4， YUAN Jiayil

（1.SchoolofArtificial Intellgence，NorthChinaUniversityofScienceandTechnology，TangshanO6321o，China; 2.Hebei Key Laboratoryof Industrial Intelligent Perception，Tangshan O6321o，China; 3.SchoolofComputerandCommunicationEnginering，UniversityofSienceandTechnologyBeijing，BeijingOoo83，China; 4.BeijingKeyLaboratoryofKnowledgeEngineeringintheFieldof Materials，Beijing1Ooo83，China）

Abstract：Inthecontext offlow velocity measurement，ripples on theriversurfacereflectthemotion characteristics of water flow，andopticalflowtimationisciticalforcapturingthismotion.However，thexistingopticalflowstimationalgoritsare primarilydesignedforrigidbodymotionandfailtofullytakeaccountofthecharacteristicsof waterflow.Toenhance the adaptabilityoftherecurental-airsfieldtransform（RAF）networkinwaterflowvelocitymeasurement，animprovedvelocity measurementmethodbasedonopticalflowestimationisproposed.Firstly，afeatureenhancementmodulebasedonmulti-scale deformableconvolutionisdesignedaccordingtothecharacteristicsof waterflowmotion，soastoenancethenetwork'sabilityto capturesimilarriplesonriversurfaces.Secondly，areal waterflowopticalflowdatasetisconstructedbasedontheRealFlow framework，andtheimagepairswithrealmotionlabelsaregeneratedbytherendering methodofrealimagepairs，soas to aleviatethedomainadaptabilityissewhenthenetworkisapliedtorealriverscenariosaftertrainedonthesyntheticrigid bodyopticalflowdataset.Finally，quantitativeevaluationofwaterflowvelocityisconductedonrealriverscenedatatovalidate thefectivenessoftheimprovedmethod.Theexperimentalresultsindicatethattheproposedmethodachievesameanrelative error （MRE） of 7.55% inactual river scenarios，demonstrating good accuracy and reliability.

Keywords：deepoptical flow estimationnetwork；flowvelocitymeasurement;RAFT;featureenhancement;realisticimage pair rendering;speed conversion model

0 引言

光流估计测速法是一种基于视频图像的非接触式流速测量方法[1-3]。（剩余9281字）