基于改进YOLOv7的密集行人检测算法  

Dense pedestrian detection algorithm based on improved YOLOv7

作　　者：黄诗佳 蒋碧波 杨超 李致君 许伶俐 HUANG Shijia;JIANG Bibo;YANG Chao;LI Zhijun;XU Lingli(School of Computer and Information Engineering,Hubei University,Wuhan 430062,China;School of Computer and Information,Hubei University,Wuhan 430062,China;Engineering Research Center of Hubei Province in Intelligent Government Affairs and Application of Artificial Intelligence,Wuhan 430062,China)

机构地区：[1]湖北大学计算机与信息工程学院,湖北武汉430062 [2]湖北大学人工智能学院,湖北武汉430062 [3]智慧政务与人工智能应用湖北省工程研究中心,湖北武汉430062

出　　处：《湖北大学学报（自然科学版）》2024年第4期522-530,共9页Journal of Hubei University：Natural Science

基　　金：国家自然科学基金(61977021);湖北省重点研发计划项目(2021BAA184)资助。

摘　　要：针对密集场景中的行人目标往往存在重叠、遮挡、体积较小等问题,导致在检测过程中容易出现漏检、误检、特征提取困难、定位不准确等现象。提出一种改进YOLOv7的密集行人检测算法,首先在YOLOv7主干网络的ELAN结构上融合CBAM注意力机制,使主干网络更加关注特征的语义信息,以增强遮挡物和小目标的特征提取能力;其次在检测头的卷积中引入CoordConv模块,充分利用此模块的位置信息,有效改善了目标定位不准确的问题,提高模型对空间位置的感受能力和泛化能力;然后将原YOLOv7网络模型中的CIoU损失函数替换为Focal-EIoU损失函数,可以有效缓解正负样本不均衡的问题,在边界框回归过程中,该损失函数更注重于高质量锚框,从而加快网络的收敛速度;最后在模型中用非极大值抑制Soft-NMS算法替换传统的NMS算法,有效降低了重叠、遮挡目标的漏检率,提升模型的召回率和精度。在公开密集行人数据集WiderPerson上进行验证本模型,实验结果表明,改进后的检测算法对密集行人目标检测的准确率、召回率及平均精度mAP值相较原基线模型分别提升了2.3%、3.3%与2.6%,FPS值提升了2.3 f/s。Aiming at the pedestrian targets in dense scenes often have problems such as overlapping,occlusion,and small size,which lead to easy leakage,misdetection,difficulty in feature extraction,and inaccurate localization in the detection process.A dense pedestrian detection algorithm to improve YOLOv7 was proposed,which firstly integrated the CBAM attention mechanism on the ELAN structure of the YOLOv7 backbone network,so that the backbone network payed more attention to the semantic information of the features,in order to enhance the feature extraction ability of the occlusions and the small targets;and secondly,introduced the CoordConv module in the convolution of the detector head,and made full use of the positional information of this module to effectively improve the problem of inaccurate target localization,and improved the model’s ability to sense spatial location and generalization;then the CIoU loss function in the original YOLOv7 network model was replaced with the Focal-EIoU loss function,which could effectively alleviate the problem of imbalance between positive and negative samples,and in the process of bounding-box regression,this loss function focused more on the high-quality anchor frames,which accelerated the convergence of the network;Finally,the traditional NMS algorithm was replaced by the non-maximum suppression Soft-NMS algorithm in the model,which effectively reduced the leakage detection rate of overlapping and occluded targets,and improved the recall and precision of the model.The validation of this model was carried out on the publicly available dense pedestrian dataset WiderPerson,and the experimental results showed that the improved detection algorithm improves the accuracy,recall and average precision mAP values of the dense pedestrian target detection by 2.3%,3.3%and 2.6%,respectively,and the FPS value was improved by 2.3 f/s compared with the original baseline model.

关 键 词：密集行人检测 YOLOv7 注意力机制 CoordConv 损失函数 非极大值抑制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]