一种抗遮挡重叠与尺度变化的行人检测算法

打开文本图片集

关键词：行人检测；人群密集场景；FasterR-CNN；多尺度特征融合；损失函数；非极大值抑制中图分类号：TN911.73-34；TP391.41 文献标识码：A 文章编号：1004-373X（2026）01-0041-08

Pedestrian detection algorithm resistant to occlusion overlap andscalevariation

MAXiming，LI Ning，WUDi

Changchun InstituteofOptics，Fine MechanicsandPhysics，ChineseAcademyofScienes，Changchun130033，China

Abstract：In view of the facts that the detectionaccuracyof pedestrian detector decreases and themissed detection rate increasesduetofactorssuchastheocclusionandthediferentscalesofpedestrians（thebjects）incomplexcrowdedscenesan improvedpedestriandetectionalgorithmthatisresistanttoocclusionoverlapandscalevariationisproposedbasedonFasterR-CNN algorithm.Inthestepoffeatureextraction，arecurrntmulti-scalefeatureextractionnetworkisdesignedtoincorporatethe atentionmechanism，whichisused tolearnmorerichanddetailedmulti-scalefeatureinformationandfocusonthekeyfeature informationtoimprovethesensitivityof thenetworktopedestrians（theobjects）ofdiferentscales.Forthelossfunctionmodule， repulsiveloss isintroducedtoreducetheinterference（causedbymutualocclusionoftheobjects）ondetection.Inthestepof postprocessing，anon-maximum suppresion （NMS）algorithmbasedonthecompensation foroverlappngrate isdesigned，sothat theactualsuppesionthresholdcanbeadaptivelyadjustedwiththechangeof thedegreeofocclusion，thusfurtherreducingthe misseddetectionofthepedestrians（theobjects）indenseplaces.Experimentalresultsshowthattheimprovedalgorithmhasbetter detectionperformance.Itsaverage detectionaccuracyontheCrowdHuman datasetand CityPersons dataset is improved by 2.5% （204号 and1.9%，respectively，and its log-averagemissrates（LAMRs） arereduced by3.5% and 3.2% ，respectively，incomparisonwith thoseofthebaselinealgorithm.Inaddition，itsLAMRofpedestrians（theobjets）ofdiferentscalesontheTJU-DHD-pedestrian datasetisalsoreduced significantly.Theproposed algorithmcanbeapplied topedestriandetectionincomplex scenes.

KeyWords：pedestrian detection; crowd scene;Faster R-CNN; multi-scale feature fusion;loss function;NMS

0 引言

行人检测技术作为计算机视觉领域中的重要研究内容，在自动驾驶、智能监控、智能机器人和人机交互等领域发挥着重要作用并具有较高的应用价值。（剩余12529字）