文献类型：期刊文献

中文题名：基于图切割和密度聚类的视频行人检测算法

英文题名：Pedestrian Detection on Videos Based on Graph Cuts and Density Clustering

作者：曾成斌 刘继乾

机构：[1]贵州理工学院电气与信息工程学院

第一机构：贵州理工学院电气与信息工程学院

年份：2017

卷号：30

期号：7

起止页码：588-597

中文期刊名：模式识别与人工智能

外文期刊名：Pattern Recognition and Artificial Intelligence

收录：CSTPCD;;Scopus;北大核心:【北大核心2014】；CSCD:【CSCD2017_2018】；

基金：贵州省自然科学基金项目(No.[2014]2081);贵州省普通高等学校创新团队基金项目(No.[2014]34)资助~~

语种：中文

中文关键词：行人检测;图切割;密度聚类;无监督学习

外文关键词：Pedestrian Detection, Graph Cuts, Density Clustering, Unsupervised Learning

摘要：现有视频行人检测方法把行人检测看成一个有监督的两类(即行人和背景)学习问题,区分视频中的行人和背景,并不能很好解决行人的姿态变化和行人间的遮挡问题.文中提出基于图切割和密度聚类的行人检测算法,把行人检测看成一个多类的无监督学习过程.在训练阶段,首先对每个训练样本计算多级梯度方向直方图-局部二分模式(HOG-LBP)特征,然后对多级HOG-LBP特征所属的每个图像块分配不同的权值.为了区别行人的不同部位并赋权值,采用基于图像块的图分割方法从背景中分割行人所在的图像块.最后,再采用基于密度峰值的聚类算法对正样本和负样本分别进行无监督的聚类.在测试阶段,首先通过计算样本特征与每个聚类中心的距离,然后使用前5个最短距离进行投票,判断其是否包含行人.实验证明,文中算法较好解决行人的姿态变化和行人间的遮挡问题,并且随着训练样本的增加,能取得和目前最优行人检测方法可比较的结果.
In the existing pedestrian detection algorithms, the pedestrian detection is considered as a supervised learning problem of two classes, pedestrian and background. Thus, the pedestrian and the background in the video are distinguished. However, the problem of variable poses and heavy occlusion can not be solved by these algorithms effectively. In this paper, a pedestrian detection algorithm based on graph cuts and density clustering is proposed. The pedestrian detection is regarded as an unsupervised learning problem of multiple classes. At the training stage, the multilevel histogram of oriented gradient-local binary pattern（HOG-LBP） features are firstly calculated for each of training samples. Then, different weights are assigned to each image block of the multilevel HOG-LBP features. To distinguish the different parts of pedestrian and assign weight, the image sample is segmented by the block-based graph cuts algorithm. Finally, the density clustering approach is used to classify the positive and negative samples into multiple cluster center respectively. At the testing stage, the distance between the multilevel HOG- LBP of test sample and every cluster center is calculated, and the five shortest distances are voted to classify the test sample. Experiments show that the proposed algorithm can handle the pose variations and partial occlusions effectively. Moreover, with the increase of training samples, the results of the proposed algorithm can be comparable to that of the state-of-the-art pedestrian detection algorithms.