对于小目标行人和密集行人的目标检测

打开文本图片集

中图分类号：TN911.73-34 文献标识码：A 文章编号：1004-373X（2026）07-0180-10

引用格式：，，，等.对于小目标行人和密集行人的目标检测[J].现代电子技术，2026，49（7）：180-189.

Abstract：Inviewofthelowtectioacuracy，oordetectionabilityforsmallobjectsanddensepedestrans，ndpooral-e detectionperformanceinteurentintelligentvicleperceptiontchologies，hispaperpropoesapedstriandetectimethodnd modellightweightmethodbasedonYOLOv8.Thebackbonenetwork isreplacedwith Swin-Transformer，andhigh-resolution featuremapsareadedtoimprovethesensitivityof themodeltothesmall targetpedestriandetection.Theimproved explicit visualcenterisembeddedintheNecklayertoenablethemodel tocapturemorelongdistancefeaturedependent information，so as toconstructachannelspaceatentionmechanismCSNA-Model （channel-spatial-none-atentionmodel），whichiscombied withcoordinateatentiontostrengthenthemodel'slearningabilityforchannelandspecialkeyinformation.Theacuracyof predictionboxgenerationpositionisimprovedonthebasisoftheimprovedlossfunctionEIoU.Theimprovedmodelistestedand analyzedontheconstructed pedestriandataset.Theresultsshowthatthedetectionaccuracyof theimprovedmodelishigherthan thatof the original YOLOv8 model. The AP @0.5 of the improved model is increased from 83.8% to 88.9% and its AP （a0.5：0.95 （20号 is increased from 51.6% to 55.2% ，so the improved model performs‘better in detecting small target pedestrians and dense pedestrians.

Keywords：pedestrian detection; smalltarget pedestrian; dense pedestrian; YOLOv8; dep learning;attention mechanism

0 引言

随着我国经济的迅速发展，汽车数量持续增长，带来了一系列与人们生活息息相关的挑战，如交通拥堵、环境安全和交通安全等问题，同时驾驶者对车辆智能化水平和便捷驾驶功能的需求也日益增长。（剩余11835字）