Journal of Computer Applications ISSN 1001-9081 2019-07-23
计算机应用 CODEN JYIIDU http://www.joca.cn
收稿日期: 2019-05-24; 修回日期:2019-06-20; 录用日期: 2019-06-24。
基金项目: 国家自然科学基金资助项目(61572085);
作者简介: 朱繁(1994—),女,江苏淮安人,硕士研究生,主要研究方向:计算机视觉;王洪元(1960—),男,江苏常州
人,教授,博士,CCF 会员,主要研究方向:计算机视觉;张继(1981—),男,江苏常州人,讲师,硕士,CCF 会员,主
要研究方向:计算机视觉。
文章编号:1001-9081(****)**-0000-00 doi:10.11772/j.issn.1001-9081.2019051051
基于改进的 Mask R-CNN 网络的行人细粒度检测算法
朱繁,王洪元
*
,张继
(常州大学 信息科学与工程学院,江苏 常州 213164)
(*通信作者电子邮箱 hywang@cczu.edu.cn)
摘 要: 针对复杂场景下行人检测效果差的问题,采用基于深度学习的目标检测中领先的研究成果,提出了一种改进的
Mask R-CNN 网络框架的行人检测算法。首先,采用 K-means 算法对行人数据集的目标框进行聚类得到合适的长宽比,通过增
加一组长宽比
(2:5)
得到 12 种 anchors 适应图像中行人的尺寸;然后,结合细粒度图像识别的技术,实现行人的高定位精度;
其次,采用全卷积网络(FCN)分割前景对象,并进行像素预测获得行人的局部掩码(上半身、下半身),实现对行人的细粒
度检测。最后,通过学习行人的局部特征获得行人的整体掩码。为了验证改进算法的有效性,将其与当前具有代表性的目标
检测方法(如 Faster R-CNN、YOLOv2、R-FCN 等)在同等数据集上进行对比。实验结果表明,改进的算法提高了行人检测
的速度和精度,并且降低了误检率。
关键词: Mask R-CNN;行人检测;K-means 算法;细粒度;全卷积网络
中图分类号:
TP391.41 文献标志码: A
Pedestrian fine-grained detection algorithm based on improved
Mask R-CNN network
ZHU Fan
, WANG Hongyuan
1*
, ZHANG Ji
(College of Information Science and Engineering, Changzhou University, Changzhou Jiangsu 213164, China)
Abstract: Aimed at the problem of poor pedestrian detection effect in complex scenes, this paper proposed an improved
pedestrian detection algorithm based on Mask R-CNN network framework which based on the leading research results in deep
learning-based object detection. Firstly, the K-means algorithm was used to cluster the object frame of the pedestrian datasets to obtain
the appropriate aspect ratio. By adding a set of aspect ratio
(2:5)
, 12 anchors could be adapted to the size of the pedestrian in the image;
Secondly, combined with the technology of fine-grained image recognition, the pedestrian's high positioning accuracy was realized;
Then, the foreground object was segmented by the full convolutional network (FCN), and pixel prediction was performed to obtain the
local mask (upper body、lower body) of the pedestrian, so as to achieve fine-grained detection of pedestrians. Finally, the overall mask
of the pedestrian is obtained by learning the local features of the pedestrian. In order to verify the effectiveness of the improved
algorithm, it was compared with the current representative object detection methods (such as Faster R-CNN、YOLOv2、R-FCN, etc.) on
the same dataset. The experimental results show that the improved algorithm improves the speed and accuracy of pedestrian detection
and reduces the false positive rate.
Keywords: Mask R-CNN; pedestrian detection; K-means algorithm; fine-grained; fully convolutional networks(FCN)
0 引言
行人检测技术由于应用的广泛性使其在计算机视觉领域
成为一个重要的分支,对视频监控、车辆辅助驾驶、智能机
器人等多个领域提供了重要的技术支持.它与行人重识别、
目标跟踪等领域的联系密切相关,被认为是一个图像检索的
子问题。
传统的行人检测方法大多以图像识别为基础,并基于人
工设计的特征提取器进行特征的提取。首先在图片上使用穷
举法选出所有物体可能出现的目标区域框,之后对这些区域
框提取 Haar
[1]
、方向梯度直方图(Histogram of Oriented
网络出版时间:2019-07-23 13:46:55
网络出版地址:http://kns.cnki.net/kcms/detail/51.1307.TP.20190723.1346.014.html