基于聚合通道特征及卷积神经网络的行人检测  

Pedestrian detection based on aggregate channel features and convolutional neural networks

作　　者：陈光喜[1] 蔡天任 黄勇[2] 王佳鑫 CHEN Guang-xi 1,CAI Tian-ren 1 ,HUANG Yong 2,WANG Jia-xin 1(1.Guangxi Key Laboratory of Intelligent Processing of Computer Images and Graphics,Guilin University ofElectronic Technology,Guilin 541004,China;2.Guangdong Engineering Technology Research Center forMathematical Educational Software,Guangzhou University,Guangzhou 510006,Chin)

机构地区：[1]桂林电子科技大学广西图像图形智能处理重点实验室,广西桂林541004 [2]广州大学广东省数学教育软件工程技术研究中心,广东广州510006

出　　处：《计算机工程与设计》2018年第7期2059-2063,2068,共6页Computer Engineering and Design

基　　金：国家自然科学基金项目(61462018);广东省数学教育软件工程技术研究中心开放基金项目(LD16124X);广西学位与研究生教育改革基金项目(JGY2014060);桂林电子科技大学研究生教育创新基金项目(2016XWYJ09)

摘　　要：为解决在复杂环境下难以尽可能多地检测到行人的问题,提出一种基于聚合通道特征、通过卷积神经网络提取特征的行人检测算法。采用聚合通道特征的算法尽可能多地产生候选框,通过卷积神经网络提取候选框内物体的深度特征,使用支持向量机分类器对候选框内的物体进行分类,检测出行人。在公开数据集Caltech和INRIA数据集上进行测试,实验结果表明,与目前主流算法比较,召回率平均提升12%,F值平均增加0.05,能有效减少计算机的计算开销。To solve the problem of detecting pedestrians in complex environment as much as possible,the pedestrian algorithm was proposed.Based on aggregate channel features,features were extracted through convolutional neural networks.Aggregate channel features were used to produce candidate boxes as many as possible,deep features of objects in these candidate boxes were extracted using convolutional neural networks.Support vector machine was used to classify objects in these candidate boxes for detecting pedestrians.The algorithm was tested on open datasets,Caltech and INRIA datasets.Experimental results show that compared with the current mainstream algorithm,the average recall rate is increased by 12% and the average F-measure is increased by 0.05 through the method,besides it can effectively reduce the computational cost of the computer.

关 键 词：行人检测 聚合通道特征 卷积神经网络 候选框 支持向量机 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]