人体行为识别的时空耦合随机图模型及其高效推理算法研究

项目名称： 人体行为识别的时空耦合随机图模型及其高效推理算法研究

项目编号： No.61503297

项目类型： 青年科学基金项目

立项/批准年度： 2016

项目学科： 其他

项目作者： 魏平

作者单位： 西安交通大学

项目金额： 22万元

中文摘要： 作为人体行为识别的关键技术，行为时空耦合建模是一项极富挑战性但未被深入研究的课题。本项目利用随机图理论，对行为的时空耦合进行系统深入的研究，提出新的概念表达、理论模型和计算方法对人体行为进行高精度高效率识别。以申请人之前的行为层次化理解工作为基础，本项目提出多属性隐结构随机图对行为时间域耦合建模、并识别具有复杂时间域耦合的行为；提出时空耦合分层随机图模型描述行为空间域耦合、并识别具有复杂物体交互的行为；对含有多个行为的长序列，提出多层次耦合随机图模型进行多行为协同分割与识别；针对传统方法求解多属性耦合图精度低、速度慢的问题，受DNA双螺旋结构启发，提出类DNA螺旋上升高效推理算法。与其他研究工作比，本项目研究的问题具有本质性、内容具有系统性，在思想概念、模型方法和学科交叉上具有创新性。本项目的研究具有重要的科学意义和应用前景，提出的模型具有良好的基础性和通用性，可应用于国计民生诸多领域。

中文关键词： 运动特性；理论建模；统计建模；行为识别；行为建模

英文摘要： As one of the key techniques in human action recognition, modeling spatial-temporal coupling of action is a challenging but non-well studied issue. This project will, systematically and deeply, investigate this issue with the stochastic graph theory. It will propose novel concepts, models, and algorithms to represent the spatial-temporal coupling and recognize human actions accurately and efficiently. Based on my previous work of hierarchical understanding of action, it proposes stochastic graphical models of multi-attributes with hidden structures to model the temporal coupling and recognize actions with complex temporal coupling. It further proposes hierarchical graphical models of spatial-temporal coupling to formulate the spatial coupling and recognize actions interacting with objects. For the long sequence with multiple actions, it proposes stochastic graphical models with multilayer coupling to simultaneously segment the sequence and recognize the actions. Inspired by the DNA double helix structure, it proposes pseudo-DNA double helix ascending algorithm to carry out the inference on the graphical models of multi-attribute coupling, which overcomes the drawbacks of low accuracy and inefficiency of conventional methods. Compared with the other work, the issues studied by this project are of essence and the contents are systematic. It is also creative in the aspects of concepts, models and cross-disciplines. This research is of great significance both in the theory and applications. For the fundamentality and generality, the proposed models can be applied to many fields of industry and human life.

英文关键词： motion property;theoretical modeling;statistical modeling;action recognition;action modeling