基于深度学习的可见光图像中行人检测方法  

Pedestrian Detection Method Based on Deep Learning in Visible Light Image

作　　者：单巍[1] 王江涛[1] 方振国[1] 崔少华[1] SHAN Wei;WANG Jiangtao;FANG Zhenguo;CUI Shaohua(College of Physics and Electronic Information,Huaibei Normal University,Huaibei 235000,Anhui,China)

机构地区：[1]淮北师范大学物理与电子信息学院,安徽淮北235000

出　　处：《武汉大学学报（理学版）》2021年第2期127-135,共9页Journal of Wuhan University:Natural Science Edition

基　　金：国家自然科学基金(61976101);安徽省高校自然科学研究项目(KJ2018B10);淮北师范大学质量工程项目(2017tykc159);淮北师范大学校企合作项目(RD/PY-01-2018,RD/ZD-01-2019,RD/YW-01-2019)。

摘　　要：相对传统的行人检测技术,基于深度学习的行人检测技术具有压倒性的优势,然而由于深度卷积网络规模庞大,需要专用的处理器,限制了行人检测系统的推广。针对上述问题,提出一种网络规模适中的深度卷积网络模型,在保证检测精度的前提下提高检测模型的普适性。以低维度的浅层卷积神经网络为基础,分别从网络层数、感受野大小和特征图3个角度出发搜索最优的网络结构,并通过有指导的实验评估确定最终的网络参数。在Daimler行人数据库上进行实验,结果表明,本文建立的网络不但网络规模适中,而且具备良好的检测性能。在Daimler、MIT、INRIA等行人数据库上进行的交叉实验验证了依本文方法建立网络的鲁棒性,表明其具有推广能力。Compared with traditional pedestrian detection technology, deep learning-based pedestrian detection technology has overwhelming advantages. However, due to the large scale of deep convolution network, the need for dedicated processors limits the promotion of pedestrian detection system. To solve this problem, this paper presents a deep convolution network model with moderate network size to improve the universality of the detection model while ensuring the detection accuracy. Based on the lowdimensional shallow convolution neural network, the optimal network structure is searched from three perspectives: the number of layers of the network, the size of the sensing field and the feature map, and the final network parameters are determined through guided experimental evaluation. On Daimler pedestrian database, the results show that the network structure designed in this paper not only has a moderate network scale, but also has a good detection performance. On Daimler, MIT and INRIA pedestrian databases, the cross-over experiments show that our network is robust and can be extended.

关 键 词：深度学习 卷积神经网络 可见光图像 检测率 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]