无监督领域自适应技术中开集关键问题研究及在视频监控中的应用

Intelligent video surveillance is an active and important research topic in computer vision and pattern recognition. One of the most challenging problem in intelligent video surveillance is distribution mismatch caused by video data captured in different surveillance cameras. Unsupervised domain adaptation can be used to reduce the distribution mismatch between the training (source domain) and testing (target domain) data. However, most of existing methods are developed based on the close-set assumption, i.e., the source and the target domain contain the same set of classes. In practice, no label information is available in the target domain, so such close-set assumption is hardly guaranteed to be valid. By breaking through the close-set limitations, this project plans to study the key research problems cased by open-set setting in unsupervised domain adaptation, i.e., the sets of classes are not the same across domains and target classes are totally unknown. This project approaches to solving the open-set problems from the following three aspects: 1) Based on the ideas of domain adversarial learning and cross-domain group-sparse decomposition, this project studies how to determine the shared classes in both domains and align the distributions of the shared classes across domains; 2) By using region metric learning to estimate label information in the target domain more accurately, this project investigates discriminative analysis between the shared classes and target-domain-specific classes; 3) With multiple source domains, based on research findings of 1) and 2), this project further studies how to model the relationship between multiple source domains, each source domain and the target domain by robust dependence modelling approach.

智能视频监控是计算机视觉和模式识别中最热门和最重要的研究领域之一。其中最具挑战性的问题之一是不同监控摄像机下采集的视频数据分布不一致。无监督领域自适应技术可用以对齐训练样本（源域）和测试样本（目标域）的联合概率。然而，多数传统方法依赖于源域和目标域所包含类别一致的闭集假设。由于实际应用中目标域没有任何标签信息，闭集假设一般难以保证。本项目将突破闭集假设局限，研究无监督领域自适应技术中开集（即领域间所包含类别不一致且目标域类别完全未知）设定下的关键科学问题。具体研究内容包括三方面：1）利用领域对抗学习和跨领域分组稀疏分解方法，研究如何确定领域间的共有类别并实现针对共有类别的联合概率对齐；2）通过区域度量学习更准确地估计目标域伪标签信息，研究目标域共有类别和独有类别的判别分析技术；3）针对多源域问题，在1）和2）的基础上，研究多源域间、源域和目标域间的鲁棒多源域关联性建模方法。

智能视频监控是计算机视觉和模式识别中最热门和最重要的研究领域之一。其中最具挑战性的问题之一是不同监控摄像机下采集的视频数据分布不一致。项目的主要研究内容是面向智能视频监控实际应用场景的无监督域适应方法，及其在自监督学习和小样本学习方面的拓展。在域适应方面，突破闭集假设下源域和目标域私有类别样本所带来的负迁移局限，提出基于对抗学习的层次特征分解网络和渐进式样本选择方法。在自监督学习方面，针对视频表征容易受静态线索影响的问题，提出对运动信息更为敏感的自监督视频表征学习方法，得到具更强泛化能力的视频表征。在小样本学习方面，针对新类与基类数据分布可能不一致且样本数极度不平衡的问题，提出基于对抗特征增强的跨域小样本分类和基于数据分布校准的交互式目标检测方法，实现小样本设定下的域间数据分布对齐和新类目标更准确的定位。项目相关的研究成果包括SCI/EI期刊论文4篇、EI会议论文17篇和中国发明专利3件，并培养10名硕士研究生顺利毕业，已完成项目任务书的目标任务。