基于Martingale-test理论的无监督人体行为分类算法研究

Vision based human action classification is a core component of practical applications of human action analysis and understanding, such as human-computer interaction (HCI), motion retrieval, vision surveillance and so on. Instead of requiring a prior supervised learning process, unsupervised human action classification is to distinguish different human actions by an unsupervised analysis of already-observed data and newly-arrived data from a view of machine learning, and hence it has a wider range of practical usage. There is, however, still a lack of universal algorithm for realizing action-video-data analysis and mining in unsupervised classification of human action. In this study, we propose to apply Martingale-test into unsupervised action classification as follows: on the basis of the Martingale-test framework, human action is first modeled as a special graph-structure by self-organizing modeling, and then Martingale-test is employed to detect the possible change point between two time-sequential graph-structures, which are respectively modeled from the already-observed action video and one newly-arrived query video, by testing the exchangeability; Last, the logical detection result, i.e., 1 or 0, is mapped into decision result(s) for action classification followed by a group of previously-established rules. This study can provide new resolutions and new theoretical cues/foundations for action-video-data analysis and mining in unsupervised action classification, and consequently achieves action recognition and new action detection for unsupervised classification of human action.

人体行为分类是计算机视觉领域中实现人体行为智能分析与理解的核心技术。无监督行为分类从机器学习的角度对行为视频进行数据分析与挖掘以区分行为类别,理论上具有更加广阔的发展前景。本课题致力于解决无监督行为分类中缺乏视频数据分析与挖掘通用算法的问题，拟提出以Martingale-test为理论框架，利用“变化(突变)检测”的方式对行为数据进行分析与挖掘，进而实现人体行为无监督分类。计算流程如下：首先从图结构角度对人体行为进行自组织建模，然后结合互换性原理利用Martingale-test实现两个时序图结构(分别由已观测行为视频和新观测行为视频构建)的变化点检测，最后将逻辑化的检测结果(1或0)映射为决策结果并结合相应处理策略实现人体行为无监督分类。本课题可为无监督行为分类提供视频数据分析与挖掘的通用算法，并建立相应处理策略与算法实现结构以实现可同时完成行为识别与新行为检测的无监督行为分类。

无监督人体行为识别是计算机视觉领域中实现人体行为智能分析与理解的核心技术之一。项目团队主要围绕人体行为识别和时间序列分析展开工作。一方面，针对人体行为识别，提出了一种局部自相似性的行为特征和一种基于DS证据理论的分类算法。该方法在低分辨率下的人体行为识别中体现了良好的识别性能。同时，为降低行为识别过程中的计算负担，分别提出了基于时间序列实时分析和回溯性分析的关键帧提取算法，并将其成功应用到视频的梗概中。本项目的研究结果对于时序数据分析在行为识别中的应用具有重要意义。. 此外，针对周期性时间序列的变化检测，分别提出了基于自回归积分滑动平均模型和图结构的检测算法，解决了周期性时间序列结构变化点检测和变化周期检测的问题，并成功应用到旋转机械运行状态监测中。这为以后研究工作的开展提供了新的研究思路。.

内容获取失败，请点击重试