项目名称: 行人检测中粒度空间特征提取方法研究
项目编号: No.61300161
项目类型: 青年科学基金项目
立项/批准年度: 2014
项目学科: 自动化技术、计算机技术
项目作者: 刘亚洲
作者单位: 南京理工大学
项目金额: 23万元
中文摘要: 本课题提出了基于粒度空间的特征提取方法用于人体检测。粒度表示特征对于数据的抽象能力,精细粒度特征对于数据有较低程度的抽象,因此具有较好的细节描述能力,适合用于对数据进行确定性的描述;而粗糙粒度的特征对于数据有较高程度的抽象,其所体现的是一种统计特性。因此,粒度空间特征意味着对于数据进行不同层次的抽象,从而得到从确定性描述到统计性描述的一系列具有不同描述特性的特征表示。在粒度空间特征框架下,我们进一步研究了特征粒度与目标尺度之间的耦合关系、多通道异质特征的快速提取方法以及利用领域自适应实现特征与人体姿态的共变。粒度空间特征的优势在于:1)具有非常丰富的描述能力,既可以提取结构特征,又可以提取统计信息;2)粒度具有明确的几何意义,可以方便的进行定量的控制;3)可以通过调节粒度参数来实现异质特征提取,算法复杂度较低。
中文关键词: 行人检测;粒度空间特征;深度学习;迁移学习;行为分析
英文摘要: A new feature extraction framework for human/pedestrian detection is presented. The concept of granularity is used to define the feature's abstraction capability for representing complex the objects. Specifically, the finely granular descriptors generate low level abstractions of the object which can represent the specific geometry information of the object; while the coarsely granular descriptors generate high level abstractions of the object which can provide the statistical representations of the object. The representation abilities of the features of different granular are complementary. By changing the granular parameter, a family of descriptors with versatile representation property can be generated. This family of representation is referred as to granular space representation of the object. Under the framework of granular space feaure, we further investigate the coupling relationship between the feature granular and object scale, multi-channel feature integration and domain adaptation based pose covariance feature. The benefits of this granular space based representation are: firstly, versatile representation ability for both the detailed structures and statistical information; secondly, the granular parameter has specific geometry meaning and can be control quantitatively; thirdly, by tuning the granular
英文关键词: Pedestrian detection;Granular Space feature;Deep learning;Transfer learning;Activity analysis