基于多尺度特征融合的行人检测方法  

Pedestrian Detection Method Based on Multi-scale Feature Fusion

作　　者：李岩 孟令军[1] LI Yan;MENG Lingjun(National Key Laboratory for Electronic Measurement Technology, North University of China, Taiyuan 030051, China)

机构地区：[1]中北大学电子测试技术国防科技重点实验室,山西太原030051

出　　处：《微型电脑应用》2021年第6期117-120,124,共5页Microcomputer Applications

摘　　要：在智能监控时,由于行人目标存在分布范围广,大小不一致的问题,影响了行人检测的效果。为了提高监控下对不同大小行人的检测精度,更好地应用于公共安全领域,对YOLOv3算法进行改进。首先在检测网络的三条分支网络引入特征金字塔池化结构,进行特征拼接实现局部与全局特征相融合,增加特征提取能力。然后在预测网络前引入自适应特征融合结构,融合不同特征图的位置与类别信息,提高特征表达能力。最后在预测网络中利用K聚类算法调整锚框的尺寸,提高模型的检测精度。结果表明对行人目标的检测精度较改进前提高了3.3%。In the intelligent monitoring,pedestrian targets have wide distribution and inconsistent sizes,which affects the effect of pedestrian detection.In order to improve the pedestrian detection accuracy of different sizes under surveillance,and get better result for the public safety,the YOLOv3 algorithm is improved.Firstly,the spatial pyramid pooling(SPP)structure is introduced into the three branch networks of the detection network,and feature splicing is performed to realize the fusion of local and global features and increase the feature extraction capability.Secondly,the adaptively spatial feature fusion(ASFF)structure is introduced before the prediction network,and the position and category information of different feature maps are merged to improve the feature expression ability.Finally,the K-means algorithm is used in the prediction network to adjust the size of the anchor to improve the detection accuracy of the model.The experimental results show that the detection accuracy of pedestrian tar-gets is increased by 3.3%compared to before improvement.

关 键 词：行人检测 特征金字塔池化 自适应特征融合 K聚类 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]