基于SSD的行人头部检测方法  

Pedestrian head detection method based on SSD

作　　者：李欢 陈先桥[1,2] 施辉[1,2] 杨英[1,2] 龚䶮[3] LI Huan;CHEN Xian-qiao;SHI Hui;YANG Ying;GONG Yan(College of Computer Science and Technology,Wuhan University of Technology,Wuhan 430063,China;Hubei Key Laboratory of Transportation Internet of Things,Wuhan University of Technology,Wuhan 430063,China;Communication and Navigation Department,China Waterborne Transport Research Institute,Beijing 100088,China)

机构地区：[1]武汉理工大学计算机科学与技术学院,湖北武汉430063 [2]武汉理工大学交通物联网技术湖北省重点实验室,湖北武汉430063 [3]交通运输部水运科学研究院电控信通中心通导室,北京100088

出　　处：《计算机工程与设计》2020年第3期827-832,共6页Computer Engineering and Design

基　　金：国家重点研发计划基金项目(2018YFC0810400)。

摘　　要：Faster R-CNN、SSD、YOLO都是针对行人整体检测,在人群密集场景检测精度低,为有效解决遮挡严重场景的行人检测问题,提出改进的SSD行人头部检测方法,使用K-means++聚类得到SSD先验框规格。针对SSD小目标检测的不足,建立改进SSD头部检测模型,利用SSD网络目标特征提取,添加类别预测和位置预测两个旁支网络实现特征分离。类别预测特征图采用上采样方式融合高语义,位置预测特征图采用下采样方式融合细节信息,融合两个预测结果得到最终目标。实验结果表明,该方法能实时准确地定位行人头部,有效地解决行人遮挡问题,提升检测精度。Faster R-CNN,SSD,YOLO aim at pedestrians’overall detection.To effectively solve the problem of pedestrian detection in crowded scenes,an improved SSD pedestrian head detection method was proposed,using K-means++clustering to get the SSD priori box specifications.To overcome the shortcomings of small target detection in SSD,an improved head detection model of SSD was established.Using SSD network target feature extraction,two side-branch networks of classification prediction and location prediction were added to realize feature separation.The classification prediction feature map was fused with high semantics by up-sampling,while the location prediction feature map was fused with the detailed information by down-sampling,the two prediction results were fused to get the final goal.Experimental results show that the proposed method can locate the pedestrian head accurately in real time,effectively solve the problem of pedestrian occlusion and improve the detection precision.

关 键 词：行人检测 头部检测 卷积神经网络 单发多盒检测器 特征分离 特征融合 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]