基于改进多尺度残差网络的行人检测方法  

Pedestrian detection based on improved multi-scale Res2NeXt

作　　者：孙佩珺 张仲荣 李琦铭 李俊[2] SUN Pei-jun;ZHANG Zhong-rong;LI Qi-ming;LI Jun+(School of Mathematics and Physics,Lanzhou Jiaotong University,Lanzhou 730000,China;Quanzhou Institute of Equipment Manufacturing Haixi Institutes,Chinese Academy of Sciences,Quanzhou 362000,China)

机构地区：[1]兰州交通大学数理学院,甘肃兰州730000 [2]中国科学院海西研究院泉州装备制造研究中心,福建泉州362000

出　　处：《计算机工程与设计》2023年第3期762-769,共8页Computer Engineering and Design

基　　金：泉州市科技计划基金项目(2020C052);甘肃省科技计划基金项目(20YF3GA013);甘肃省科技型中小企业技术创新基金项目(20CX9JA128)。

摘　　要：针对行人检测的尺度变化问题,提出一种基于改进多尺度残差网络无锚检测算法(IMSNet)。将Res2Net残差模块中多尺度特征提取融入ResNeXt,将改进后的网络作为主干网络(Res2NeXt*),使主干网络包含不同数量、不同组合的感受野;利用多个较小卷积核等效替代单个较大卷积核,增加网络深度并减少网络参数量;对细化的多尺度卷积特征级联融合做卷积运算,将行人检测简化为中心点和尺度预测任务。实验结果表明,IMSNet对CityPersons和Caltech数据集Reasonable设置分别实现了10.6%和2.6%的平均漏检率,检测每张图像仅需0.28 s。Aiming at the scale variation problem of pedestrian detection,an anchor-free detection algorithm based on improved multi-scale Res2NeXt(IMSNet)was proposed.Multi-scale feature extraction from the Res2 Net residual module was integrated into ResNeXt,and the improved network was used as the backbone(Res2NeXt*),making the backbone contain different numbers and different combinations of receptive fields.Multiple smaller convolution kernels were used to equivalently replace a single larger convolution kernel,the depth of the network was increased and the amount of network parameters was reduced.Pedestrian detection was simplified as straightforward central point and scale prediction task by performing convolution on the concatenation of refined multi-scale convolution features.Experimental results show that with reasonable setting of CityPersons and Caltech dataset,the IMSNet can achieve an average miss rate of 10.6%and 2.6%,respectively,and it only takes 0.28 s to detect each image.

关 键 词：机器视觉 行人检测 深度学习 尺度变化 多尺度残差网络 特征融合 级联融合 感受野 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]