基于改进YOLOv8+DeepSORT的多行人追踪算法

打开文本图片集

中图分类号：TP391 文献标志码：A DOI：10.3969/j.issn.1673-3819.2026.02.011

Abstract：Toaddresstheissueof trajectory interuptionscausedbytargetoclusionandappearance similarityinmulti-pedestrian tracking withincomplex scenes，this study proposes arobust tracking algorithmthatintegratesanimproved YOLOv8 withDeepSORT.Methodologically，thedetectionnetwork backbone isreconstructedusingthe InceptionNext Block toenhancefeaturerepresentationcapabity，and the VoV-GSCSP module is employed tooptimize multi-scale feature fusion.Additionally，themotionpredictionmodelisimprovedbyincorporating theNSAKalmanfilter，andtheAFLinkalgorithmisintroduced toachievecross-frametrajectoryassociation.Experimentalresultsdemonstratethattheimprovedmodelachievesa 0.9 increase in MOTPanda0.7increase in HOTAonthe MOT2Odataset，significantlyenhancing localization accuracyand trajectory continuityinocclusionscenarios.This providesan effective technical solution for dense crowd tracking.

Key Words：pedestrianre-identification；object detection；multi target tracking；YOLOv8；DeepSORT

随着计算机视觉技术的发展，多行人追踪在智能监控等领域的重要性日益凸显，但目标遮挡、外观相似性等挑战仍制约其性能[14]。（剩余10253字）