面向密集人群场景的多径聚合姿态估计网络研究

打开文本图片集

关键词：多人人体姿态估计；YOLOv11；关键点检测；多尺度融合；小目标检测；深度学习中图分类号：TN911-34；TP391.41 文献标识码：A 文章编号：1004-373X（2026）09-0091-07

Multi-path aggregation pose estimation network for dense crowd scenarios

MiTaihang，Guo Chenxia，Yang Ruifeng （School of Instrumentand Electronics，North UniversityofChina，Taiyuan O3oo51，China）

Abstract：In view ofthechalengesof diffcult feature extractionandlowrecognitionaccuracyofsmallobjectsinhuman poseestimation indensecrowdscenarios，thispaperdesignsamulti-pathaggegate block（MPA-Block）.Itusesmulti-branch convolutiontosynergisticallyenhancethefeatureextractionabilityindensecrowdbackground，whichimprovesthe representationdimensionofmulti-levelfeaturesandavoidstheinsuficient featureextractionabilityofC3k2intheoriginal framework.Amulti-scaledilatedaentionblock（MDA-Block）isdesigned，whichmodifiesthespatialatention（C2PA）module viaamulti-scaledilatedatention（MDA）mechanism.Thismodification increases thenumberofselectable dilationrates，nabling themodeltodaptivelyfocusonfne-grainedlocalfeatures.Specificalytwodilationratesarededicatedtosmallobects，hreby improvingtherecognitionaccuracyofsmallobjectsindensescenarios.TheYOLOseriesalgorithmsareadoptedasthebasic framework，andaftercomprehensiveexperimentalcomparisons，YOLOv1withoptimalperformanceisselectedasthebncark model.Accordingly，amulti-scaleaggregationposenetwork （MSA-PoseNet）tailoredforcomplexscenariosisdeveloped. Experimental resultson the dataset CrowdPose demonstrate thatthe proposed methodachieves77.7%and47.7% in mAP @0.5 2 andmAP@0.5：0.95，respectively.Incomparisonwiththeoriginalmodel YOLOv11，theproposedmethodachievesan improvement of 2% and 2.3% in mAP@0.5 and mAP@0.5：0.95，which validates the effectiveness of the proposed method in the task ofcomplex crowd pose estimation.

Keywords：multi-person human pose estimation; YOLOv11; key point detection; multi-scale fusion;smallobject detection; deep learning

0 引言

视觉技术也不断发展，其理论研究和应用的方向也变得越来越广泛。（剩余11044字）