基于深度表达和迁移学习的人体检测研究

项目名称： 基于深度表达和迁移学习的人体检测研究

项目编号： No.61502173

项目类型： 青年科学基金项目

立项/批准年度： 2016

项目学科： 计算机科学学科

项目作者： 吴斯

作者单位： 华南理工大学

项目金额： 20万元

中文摘要： 随着视觉监控的广泛应用，对监控系统的智能化需求越来越迫切。对于监控对象是人的场景，人体出现的时间和位置是最关键的信息，而实现自动人体检测是获取这些信息的必要手段。对人体检测的研究在过去的20年有很大进展，但在特征学习和模型领域适应性方面仍有很大改进的空间。现有特征提取方法难以挖掘图像高层次信息，本课题拟采用深度卷积网络学习图像多层次表达，并且通过随机池化和基于稀疏支持向量机的特征选择算法获取多层次特征。我们拟提出基于级联模型的检测器，其中节点与多层次特征逐个关联。由于监控场景千差万别，源域与目标域数据分布之间可能存在很大差异，本项目拟采用迁移学习方法解决领域适应性问题。在给定少量目标域样本的情况下，我们拟采用基于模型的迁移学习策略，即通过预训练检测器稀疏组合构建新检测器，从而保证新检测器泛化能力强。本课题的开展将会为提升人体检测性能和增强其实用性提供新思路。

中文关键词： 人体检测；特征选择；迁移学习

英文摘要： With wide applications of visual surveillance, the need for intelligent monitoring system becomes very urgent. For the scenarios in which the monitoring object is human, the critical information is the time and place of human appearing, and human detection is an important step to obtain that. Great progress has been made in human detection in the past 20 years. For feature extraction and domain adaptation, many successful methods have been reported, though there is still room for improvement. In order to extract discriminative features for our task, we expect to employ deep convolutional neural network to learn the multi-level representation of images, stochastic spatial pooling to generate candidate features, and sparse support vector machine to select the important components from them. On the other hand, the difference between the distributions of source and target domain data may be significant because of the variety of application scenarios. As a result, domain adaptation is needed to solve this problem. Given a small number of labeling samples in the target domain, we propose to develop a cascaded classifier-based detector for which a transfer learning approach is used to build the nodes sequentially. Specifically, each node of a new cascade is constructed sparsely by the corresponding nodes of a set of pre-trained cascades such that the new detector has high generalization ability. This project will provide a promising way for human detection to boost its performance and improve its applicability.

英文关键词： Human Detection;Feature Selection;Transfer Learning