计算机与现代化

• 图像处理 • 上一篇    下一篇

 基于广义霍夫变换的室外场景行人检测研究

  

  1. 大连交通大学软件学院,辽宁大连116052
  • 收稿日期:2014-12-22 出版日期:2015-04-27 发布日期:2015-04-29
  • 作者简介:张雪松(1980-),男,湖北襄樊人,大连交通大学软件学院讲师,硕士,研究方向:机器学习与模式识别。

 Research on Human Detection in Outdoor Scene Based on Generalized Hough Transform

  1. Software Technology Institute of Dalian Jiaotong University, Dalian 116052, China
  • Received:2014-12-22 Online:2015-04-27 Published:2015-04-29

摘要:

 提出一种基于广义霍夫变换的室外场景行人检测方法。首先从少量标注图片中随机地提取行人图像碎片构造碎片字典,然后使用图像碎片对每一幅训练图片计算特征向量。为了能够在静态
图片中快速地检测行人,使用Gentleboost算法训练检测器,在每一次迭代时学习一个决策树桩弱分类器,该弱分类器可以从高维特征向量中选择一个当前区分度最好的碎片特征。在运行检测器时,所有
的弱分类器在测试图片中对于行人的可能出现位置进行投票。最后,将各个弱分类器的投票结果进行叠加,并用设定的检测阈值剔除得分较低的检测结果后得到检测输出。在LabelMe数据集上的实验表明
,该方法可以快速地在静态图片中检测出行人,需要较少的训练数据且有效地解决了部分遮挡问题。

关键词:  , 行人检测, 广义霍夫变换, 碎片特征, 室外场景

Abstract:

 A novel outdoor scene human detection approach based on generalized Hough transform is proposed in this paper. Firstly, we randomly extract fragments of human
instances from annotated training images and build a fragment dictionary. Then we utilize these extracted fragments to compute feature vectors for each training image. In order
to rapidly detect human in a static image, Gentleboost algorithm is used to train detectors. In each round of boosting, a regression decision stump is learned as the weak
classifier which can pick the most distinctive fragment features from the high dimensional fragment feature vector. When running the trained human detector, all the weak
classifiers voted for the possible positions of human instances in a given test image. Finally, the output positions of all the weak classifiers are accumulated and some low
score outputs are eliminated using a manually specified threshold to get the final detection outputs. Experiments on LabelMe datasets show that this approach can rapidly detect
human instances from static image using relative fewer training images and can effectively solve the partial occlusion problem.

Key words:  human detection, generalized Hough transform, fragment features, outdoor scene