YOLOv5s的改进算法及其在道路车辆行人检测的应用研究

打开文本图片集

中图分类号：TP274 文献标志码：B 文章编号：1671-5276（2025）06-0184-06

Abstract：Toovercomethetechnicaldifcultiessuchasthedificultyfpedestriandetection，longtermsmalltargetandmultiscaletargetintheinteligentvehicleofautomobiles，amulti-scalevehiclepedestriandetectionandrecognitionmodelYOLOv5sCDintheroadbackgroundwasproposedbasedontheYOLOv5snetwork，whichusedthepriorknowledgeoftheConvNeXtnetwork andtheperformanceadvantagesof PConvconvolutiontoimprovetheoriginalbackbonenetworkandenhancethefeatureextraction performance.Channel weightsare integratedintothedeformableconvolutionthatcanadaptto featurechanges，amulti-scale cascademodule isconstructedbasedonmore atentiontoimportant spatialregionsand channels，andtheshallowfeature layer is usedtoimprovetheextractionperformanceofsmall targetfeatures.Experimentsshowthat intheBDDlOoKdataset，the recognition accuracy of the improved YOLOv5s-CD model reaches 96.4% ，the recall rate reaches 87% ，with average accuracy as （204号 75% .Compared withtheoriginal network andother mainstreamnetwork models，recognitionaccuracyand recallrate are both improved.

keywords：vehiclepedestrian detection；YOLOv5s；ConvNext；variabilityconvolution

0 引言

精准高效的车辆行人检测技术是智能驾驶环境感知当中重要的任务之一，伴随着传感器和计算机技术的快速发展，车辆行人检测技术也在飞速进步之中[]。（剩余7911字）