全局上下文引导的双代价聚合立体匹配网络

打开文本图片集

关键词：立体匹配；特征提取；全局上下文信息；代价体；双分支代价聚合；多尺度特征；特征融合；视差回归中图分类号：TN711-34；TP391.41 文献标识码：A 文章编号：1004-373X（2025）17-0104-08

Stereo matching network with global context guided dual-cost aggregation

FANG Weizhou，MENG Xiaoyan1，2，3，ZHOU Hong1 ，DING Xiaochen1 （1.SchoolofComputerandInformationEngineering，XinjiangAgriculturalUniversity，Urumqi83oo52，China; 2.Ministry of Education Engineering Research Center for Intelligent Agriculture，Urumqi 830o52， China; 3.XinjiangAgricultural Informatization Engineering TechnologyResearch Center，Urumqi 83oo52，China）

Abstract：However，currentstereomatchingalgorithmssillfacesubstantialchallengesintexturelessareas，occluded regions，andareaswithblurrededge.Inviewofthis，aglobalcontextguideddual-costaggregationstereomatchingnetwork GCDANetisproposed.Firstly，inthephaseoffeatureextraction，theatentionmodulethatguidestheglobalcontextfeaturesis usedtocapturethedetailsoffeaturesandrichglobalcontextinformation，soastoimprovetheexpresiveabilityandrousess offeatures.Next，agroup-wisecostvolumeandaConcatVolumeareconstructed，with each cost volume procesedseparately. Then，adual-branchcostaggregationstructureisproposed.Withthedesignedmulti-scaleatentionfeaturefusionmodule，the featuresoftheupperandlowerbranchesof thecostaggregation networkandthemulti-scale featuresinthefeatureextraction modulearefsed，sotoaicuatedigutiogoetricfoaionallspritesisldto obtainthedisparitymap.TheexperimentalresultsshowthattheperformanceofGCDANetisbeterthanthatofthebenchmark model GWCNet on multiple datasets.Specifically，the evaluation indicator EPE and D1 on the SceneFlow dataset are reduced to 0.60 pixel and 2.08% ，respectively;onthe KITTI2O12dataset，the evaluation indicator3pixel-All and3 pixel-Nocarelowered to 1.61% and 1.29% ，respectively;andon theKITTI2O15dataset，theevaluation indicator D1 -All forall pixel regionsisreduced to 1.94% .The proposed network has strong adaptabilityand accuracy in handling complex scenes，and has a wide application prospect in thefieldsthatneed precise stereo matching，such asautonomous driving，objectdetectionand3Dreconstruction.

Keywords：stereo matching;featureextraction;globalcontextinformation;costvolume;dual-branchcostaggregation; multiscale feature; feature fusion; disparity regression

0 引言

双目立体视觉是计算机视觉领域中一个重要的研究方向，该技术基于视差原理恢复物体三维几何信息，在三维重建、机器人、无人驾驶、目标检测等-4诸多领域具有广泛应用。（剩余12111字）