面向复杂场景的群体活动多粒度解析研究

随着互联网和视频监控技术的发展，视频成为目前传播信息的主要载体之一，并呈现爆炸式增长。这些视频中很大一部分内容包含着多人参与的复杂人群活动。本项目以群体活动场景中的“人”为研究对象，对群体活动中的个体行为、交互行为、以及群体行为开展多粒度式的分析与理解研究，并可视化地构建与模拟人体行为的演变模式，力求达到复杂场景下群体活动的透彻性解析。拟重点研究的内容包括：1）群体活动场景下个体行为、交互行为与群体行为的特征学习；2）群体场景中个体行为、交互行为与群体行为等多粒度行为的联合识别；3）基于早期行为的群体整体行为预测；4）基于人体骨骼点序列与视频序列的人体未来行为生成。最后，本项目拟搭建一套智能化的群体活动解析应用系统，对项目研究的各种方法进行测试和性能评价。本课题的研究方法可为群体活动的智能分析与理解提供有力的核心技术，其研究成果能为完善平安城市、天网工程的建设产生积极意义。

With the developments of Internet and video surveillance technologies, video becomes one of the main carriers for spreading information, and its number is explosively increasing. Most of these videos describe the complex group activities joined by multiple persons. In order to thoroughly understand group activity in complex scenario, we focus on “human” in group activity scenes, study the multi-granularity actions, i.e., individual action, human interaction, and group action, visually build the temporal motion dynamic of human motions. In this proposal, we target to address the following key issues：1) representation learning of individual action, human interaction, and group action in group activity scenes; 2) joint recognition of the multi-granularity actions including individual action, human interaction, and group activity in group activity scenes; 3) group activity prediction based on a part of earlier activity; 4) human future action generation in human skeleton-motion sequence or video sequence. Finally, this proposal will develop an intelligent group activity understanding system for evaluating and validating the performance of the proposed methods. The proposed methods could provide the key technologies for intelligent analysis and understanding of the group activity, and further assist with the development of the Safe City and Skynet Project.