交互行为的自动理解在智能监控、智能人机交互等领域有着广泛的应用前景，是计算机视觉领域备受关注的新方向。现有的交互行为理解方法存在两个问题，一是交互行为模型主要是通过直接抽取大量的视频特征建立的，通常不能准确反映交互行为的内在规律；二是大多数交互行为识别方法是直接针对运动信息进行分析，没有将底层的人体部件与高层的环境上下文信息进行有机结合。针对这些问题，本项目主要研究视频中交互行为的内在规律，基于概率统计模型并结合上下文信息，建立交互行为自动识别与理解的理论与方法。内容包括：研究行为视频单词或行为基元的码本学习算法，建立个体行为之间的关系描述子"交互短语"，研究基于层级条件随机场模型的交互行为识别方法，以及基于环境上下文的交互行为理解方法。为了验证所提出的理论和算法，本项目还建立并发布交互行为视频数据集。

The analysis of interactive actions is a hot topic in the domain of computer vision and has promising applications in intelligent video surveillance, content based video retrevial, intelligent human-computer interaction and etc.At present, most interactive activity models are built by first extracting a lot of local features, then dimension reduction and clustering techniques are used to obtain the codebook of activity, the resulted model usually cannot express the inner feature of the activity accurately. Alos most of current activity analysis method focus on activity recognition itself, and do not combine the lower features and context infromation together to fully utilize the information contained in the interactive activity. In this project, we will focus on the inner features of interactive activity to establish effictive activity model for interactive activity understanding. The research include: 1) Establish the model of individual activity by Weighted Adaptive Metric Learning, and establish the model of interactive activity by interactive phase; 2) Combine the lower level features (human part), activity recognition, and the context infromation together to recognize the interactive activity effectively. 3) Collect and publish video data set of interactive activity.