单目移动拍摄下基于隐式形状模型的行人检测方法研究

项目名称： 单目移动拍摄下基于隐式形状模型的行人检测方法研究

项目编号： No.60873179

项目类型： 面上项目

立项/批准年度： 2009

项目学科： 轻工业、手工业

项目作者： 李绍滋

作者单位： 厦门大学

项目金额： 35万元

中文摘要： 项目围绕着单目移动拍摄下的行人检测技术展开研究，形成了以下五个方面的研究成果。首先我们建立了真实场景下的厦门大学行人数据库，针对直方图特征无法刻画象素的空间分布情况和维数较高这一缺点，提出了基于结构化局部统计特征的描述子；针对隐式形状模型中的码本构建，提出了分聚分建的类别可扩充场景分类方法和基于码本权重的特征选择算法；提出了基于迁移学习的半监督行人分类方法；提出了滑动窗口下基于直方图交叉核的多特征融合行人检测方法。另外，在本课题的研究过程中，从隐式形状模型这一思路出发，开展了基于码本的图像标注方法研究；在这一方面的研究中，也取得了较好的研究成果，已发表相应的SCI文章1篇，其中的主要成果包括基于区域空间与词汇加权的图像自动标注、基于判别模型与生成模型的层叠图像自动标注等。总之，通过本项目的研究，我们总共发表国际期刊论文SCI检索4篇、EI收录27篇、待出版专著1本，部分后期的成果还在整理中，完成了项目预期研究成果的目标，并拓宽了原有的研究内容。

中文关键词： 行人检测技术；码本构建；多特征融合；迁移学习；

英文摘要： This project focuses on the detection of pedestrian in monocular images, yields four novel fruits as follows. (1) A real-scene pedestrian dataset was built, and a structured local edge pattern feature was proposed to encode the spatial and edge information of pedestrian image; (2) For the codebook construction in the implicit shape model framework, a scene categorization with efficient class extendibility and effective discriminative ability is proposed. An information gain and SVM based visual word selection method is also proposed, which improves the performance of efficient subwindow search for pedestrian detection. (3) A semi-supervised pedestrian classification method based on transfer learning was proposed; (4) A multiple features combined method based on Histogram Intersection Kernel (HIK) Support Vector Machine (SVM) for pedestrian detection is proposed to make the detector more robust to clutter background, partial occlusion. In the research process of this project, we also extend the research contents, especially on the image annotation and human action recognition. Some image annotation methods, such as region and weighted word based image annotation, discriminative and generative model based image annotation, are proposed. Based on this project, we have pulised 4 SCI indexed papers and 27 EI index paper. There are also a book about pedestrian detection, some research paper to be published. All these achivements have significant academic and scientific values. In addition the project proves to have promising practical applications too.

英文关键词： Pedestrian detection technology; codebook construction; multiple feature fusion; transfer learning;