改进Y0L0v5下的行人车辆多目标识别跟踪检测

打开文本图片集

关键词：多目标识别；YOLOv5；MobileNetV3；深度可分离卷积；CA注意力机制中图分类号：TP391.4 文献标识码：A 文章编号：2096-4706（2025）17-0073-06

Abstract：The multi-object recognitionand trackingof pedestrians and vehicles incomplex traffc environments areof significantimportanceforensuringroad safetyandoptimizing trafficflow.Problems intraffc scenarios include large scale variations oftaretssevereolsios，andighdemandsfortimeline.Toaddressteseissues，tispaperpropoesanioved YOLOv5based on MobileNetV3 formulti-objectrecognition and tracking detection of pedestrians and vehicles.Byreplacing theC3module inYOLOv5with the lightweightMobileNetV3 network，thecomputational speedofthe model issignificantly enhanced.TheuseofDepthwise SeparableConvolutions insteadof taditionalconvolutions improves theperformanceoffeature extraction anddetection speed.The introduced CA Mechanism further optimizes the feature fusionprocess，enhancing the model'sabilitytorecognizesmallandoccudedtargets.Experimentalresultsshowthattheimprovedmodelachievesincreases of 6.7% 8.4% and 2.2% in AP， Recall， and mAP0.5 respectively compared to the original YOLOv5，and demonstrates higher confidence and better detection performance incomparison with the FasterR-CNN algorithm.

KeyWords： multi-object recognition; YOLOv5; MobileNetV3; Depthwise Separable Convolution; CA Mechanism

0 引言

随着城市化的不断发展和汽车数量的快速增加，交通场景愈加多样化，行车环境日益复杂化，因此能够高效准确检测出道路场景信息尤为重要1]。（剩余7120字）