[1]何 薇,陈宇拓.景区行人检测 YOLOv5-GSPE 算法模型研究与实现[J].计算机技术与发展,2023,33(09):113-118.[doi:10. 3969 / j. issn. 1673-629X. 2023. 09. 017]
　HE Wei,CHEN Yu-tuo.Research and Implementation of YOLOv5-GSPE for Pedestrian Detection in Scenic Spots[J].,2023,33(09):113-118.[doi:10. 3969 / j. issn. 1673-629X. 2023. 09. 017]

点击复制

景区行人检测 YOLOv5-GSPE 算法模型研究与实现()

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:

33

期数:

2023年09期

页码:

113-118

栏目:

人工智能

出版日期:

2023-09-10

文章信息/Info

Title:

Research and Implementation of YOLOv5-GSPE for Pedestrian Detection in Scenic Spots

文章编号:

1673-629X(2023)09-0113-06

作者:

何 薇; 陈宇拓

中南林业科技大学 计算机与信息工程学院,湖南 长沙 410004

Author(s):

HE Wei; CHEN Yu-tuo

School of Computer & Information Engineering,Central South University of Forestry and Technology,Changsha 410004,China

关键词:

行人检测; YOLOv5-GSPE; GhostConv; PrFPN; EioU

Keywords:

pedestrian detection; YOLOv5-GSPE; GhostConv; PrFPN; EioU

分类号:

TP181

DOI:

10. 3969 / j. issn. 1673-629X. 2023. 09. 017

摘要:

针对景区内高密度行人检测中遮挡与小目标行人漏检率高、模型复杂度高、计算量大的问题,提出一种 YOLOv5 -GSPE 改进算法模型,在保证精度的同时改善检测效果,

降低模型复杂度。 改进算法模型通过 GhostConv 优化主干网络中常规卷积( Conv)降低模型复杂度,并使用空洞卷积改善 SPPF 模块中池化操作带来的特征信息丢失,提升模型检测时效性,增强主干网络特征提取。 提出一种增强的特征金字塔网络—PrFPN,使用同层连接进一步丰富原始输入特征的融合,减少特征提取过程中的特征损失。 将引入正态分布计算优化后的 EIoU 损失函数作为边界框回归损失函数,提高边界框定位精度。 实验结果表明,YOLOv5-GSPE 算法模型对比 YOLOv5s 模型在保证检测时效性的情况下整体复杂度降低了12. 51% ,基于 Pedestrian 测试集的平均精度提升 4. 05% ,基于 WiderPerson 测试集的平均精度提升 3. 28% ,并降低了行人遮挡及小目标漏检率,改善了检测效果,该模型的可行性与有效性得到验证。

Abstract:

Aiming at the problems of high missed detection rate of occlusion and small target pedestrians,high model complexity and largeamount of calculation in high-density pedestrian detection in scenic spots,an improved YOLOv5 -GSPE is proposed,which can ensurethe accuracy while improving the detection effect. The improved algorithm model reduces the complexity of the model by optimizing theconventional convolution ( Conv) in the backbone network through GhostConv,and uses atrous convolution to improve the loss of featureinformation caused by the pooling operation in the SPPF module,improving the timeliness of model detection and enhancing the featureextraction of the backbone network. An enhanced feature pyramid network,PrFPN,is proposed to further enrich the fusion of originalinput features using same-layer connections. The optimized EIoU loss function is used as the bounding box regression loss function toimprove the positioning accuracy of the bounding box,and the proposed method educes feature loss during feature extraction. The experimental results show that compared with the YOLOv5s model,the complexity?

of the YOLOv5-GSPE is reduced by 12. 51% ,the averageaccuracy based on the Pedestrian test set is increased by 4. 05% ,and the average accuracy based on the WiderPerson test set is increasedby 3. 28% ,and it reduces pedestrian occlusion and small target missed detection. The feasibility and effectiveness of the model have beenverified.

常用功能