多尺度自注意力和局部匹配的光流估计方法

打开文本图片集

中图分类号：TP391.4 文献标识码：A DO1： 10. 7535/hbgykj. 2025yx05002

Abstract：Toadress the issues of limitedreceptive fieldandedge blurring inoptical flow estimation，anoptical flow estimation model basedon multi-scaleself-atentionandlocalfeature matching was proposed.This modelwasanimprovement uponthe recurrental-pairsfield transforms （RAFT）model.Firstly，amulti-scale self-attentionmechanism was integrated into thefeature extraction module，which learned thedependenciesbetweenlong-distance pixels using multi-scaleself-atention to obtain image feature information.Secondly，a local matching module was added during the upsampling processof low-level opticalflowtogenerate highresolutionopticalflow.Then，themodel wastrainedonopticalflowestimationdatasets.Finally， ablation experiments and comparative experiments wereconductedon the trained model.Theresults show thatthe proposed model achieves average end point eror（AEPE）of1.18and1.67on the MPI Sintel Clean and MPI Sintel Final datasets， respectively，and 1. Ol and 3.40% for average end point error and flow error of all （Fl-all） on the KITTI-2Ol5 dataset，all outperforming RAFT.The proposed opticalflow estimation model exhibits highacuracy inopticalflow estimation，which can provide effective support for computer vision tasks relying on high-precision motion information.

Keywords：computer neural network；optical flow estimation；atention mechanism；upsampling；local matching

光流形式的运动是人类视觉系统捕获运动信息的主要线索之一，能够显著增强人类的视觉认知、结构理解及自我定位能力。（剩余16193字）