频域空间信息驱动的特征聚合跨模态行人重识别方法

打开文本图片集

关键词：跨模态；行人重识别；数据增强；频域空间信息；特征聚合

中图分类号：TP391.41 文献标志码：A 文章编号：1001-3695（2026）01-036-0298-07

doi：10.19734/j.issn.1001-3695.2025.04.0143

Feature aggregation cross-modality person re-identification method driven by frequency domain spatial information

Jin Jing，Zhu Chuanbin†，Zhai Fengwen （SchoolofElectronicand InformationEngineering，Lanzhou Jiaotong University，Lanzhou 73oo7O，China）

Abstract：Thecro-modalitypersonre-identificationaimstomatchpersonimagesunderdiferentmodalitiesofvisibleandinfrared.Thecorechallngeof thistaskis toallviatethediferencesbetweenvisibleandinfraredimagesandextractdiscriminativesharedfeaturesHowever，existing methods fail tofullutilizemodalityinformationafterdataaugmentationandoverlook thesemanticcorrelationbetweenfeaturesatdiferentscaleswhileminimizingmodalitydiffrencesandextractingmodality sharedfeatures.Thispaperproposedafrequencydomainspatialinformationfeatureaggregation（FDSIFA）network.Firstly，it designedamulti-branch frequency-spatialperception module（MFSPM）tofullextractmodality-specificinformation fromboth augmentedandoriginalimages，whileexploredcros-modalityfeatureconsistencyinboth frequencyandspatialdimensions， effectivelyreducedmodalitydiferences.Then，itdesignedamulti-stagefeatureaggregationmodule（MFAM）toadaptively fuse features atdiferentscales，exploredthesemanticrelationshipsbetweenlow-levelandhigh-levelfeatures，andenhanced semantic representation and discriminability. The proposed network achieved rank-1 accuracy of 75.09% and mAP of 71. 35% intheall-search modeonthe SYSU-MMo1 dataset，outperforming existingcomparison methods.Theexperimentalresultsconfirm the effectiveness of the proposed approach.

Key words：cross-modality；personre-identification；dataaugmentation;frequencydomain spatial information；featureaggregation

0 引言

随着人工智能赋能的智慧监控系统的快速发展，行人重识别技术在智能安防领域得到广泛运用，行人重识别[12]是指通过在多个非重叠摄像头拍摄的图库中检索出特定行人的技术。（剩余16818字）