基于改进Faster R-CNN的近红外夜间行人检测方法  

Near IR night pedestrian detection method based on improved Faster R-CNN

作　　者：胡均平[1] 孙希 HU Junping;SUN Xi(College of Mechanical and Electrical Engineering,Central South University,Changsha 410000,China)

机构地区：[1]中南大学机电工程学院,湖南长沙410000

出　　处：《传感器与微系统》2021年第8期126-129,共4页Transducer and Microsystem Technologies

基　　金：国家自然科学基金资助项目(51175518);湖南省科技型中小企业技术创新基金资助项目(12C26214305029)。

摘　　要：为了解决近红外夜间行人检测中传统行人检测方法识别速度慢、准确率低的问题,提出了一种改进的快速区域卷积神经网络(Faster R-CNN)夜间近红外行人检测方法。为了提取更深层次的图像信息,将特征提取网络设计为ResNet-50;统计目标的纵横比,引入附带目标比例的加权锚点框机制;对多层次的特征图进行特征图融合。在自建夜间近红外行人数据集上的实验结果表明:使用的方法与传统LBP+Haar+HOG+SVM算法及经典Faster R-CNN算法相比,夜间近红外行人检测的平均准确率(mAP)分别提升了50%和11.2%,检测速度从2 000 ms/f和214 ms/f提高到190 ms/f。In order to solve the problem of slow recognition speed ow accuracy of traditional pedestrian detection methods in near-infrared(IR)night pedestrian detection,an improved Faster R-CNN night-IR pedestrian detection method is proposed.In order to extract deeper image information,feature extraction network is designed as ResNet-50.Aspect ratio of target is calculated,and the weighted anchor frame mechanism with target scale is introduced.Feature map fusion is performed on multi-level feature map.The experimental results on the self-built night near-IR pedestrian dataset show that compared with the traditional LBP+Haar+HOG+SVM algorithm and the classic Faster R-CNN algorithm,mean average precision(MAP)of night near-IR pedestrian detection is increased by 50%and 11.2%,and the detection speed is increased from 2000 ms/f and 214 ms/f to 190 ms/f.

关 键 词：夜间行人检测 近红外图像 快速区域卷积神经网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]