基于集成流形学习的监控视频中人体行为识别研究

项目名称： 基于集成流形学习的监控视频中人体行为识别研究

项目编号： No.61462035

项目类型： 地区科学基金项目

立项/批准年度： 2015

项目学科： 计算机科学学科

项目作者： 罗会兰

作者单位： 江西理工大学

项目金额： 46万元

中文摘要： 人体是结构复杂的非刚体，人体运动是一个复杂的运动系统；此外，运动节奏和风格的差异性，杂乱背景，遮挡和光照变化等，都使得运动人体行为识别具有非常大的挑战性。人体运动分析研究具有广泛的应用前景，在智能视频监控，医疗研究和人机交互等领域中有着巨大影响。本项目拟运用流形学习和集成学习技术对监控视频中运动人体行为识别进行研究，提出基于人体运动流形基元库，使用基本的、简单的和局部的运动流形元素来表达和构造复杂行为的思想。首先研究视频空间行为表示到流形空间的映射模型，然后研究运用集成学习技术得到人体运动流形基元库的方法。在此基础上，研究基于人体运动流形基元库的行为表示方法。最后研究基于集成学习与人体运动流形基元库的，能融合多方面信息的人体行为识别算法。建立基于视频监控系统的异常人体行为识别系统，为理论研究成果提供应用和验证平台。项目的研究将进一步完善运动人体行为识别研究的理论与应用。

中文关键词： 人体行为识别；动作识别；流形学习；集成学习

英文摘要： The human body is non-rigid with complex structure. Human motion is a complex motion system. In addition, behavior recognition is very challenging because of the pace and style diversity of human motions, messy background, occlusion and illumination variance. Human motion analysis has broad application prospects, and has a huge impact in lots of areas, such as intelligent video surveillance, medical research and human-computer interaction and so on. This project aims to adapt manifold learning and ensemble learning for human behavior analysis. The main idea is to use basic, simple and local motion manifold structure elements, which form a library of human motion manifold elements, to express complex human behaviors. First extraction methods of low-dimensional manifolds embedded in high-dimensional human motion structure will be studied. Then the ensemble learning method to obtain the dictionary of human motion manifold elements will be researched. The description of human behavior using motion manifold elements will be researched. At last, the human behavior analysis based on ensemble learning and the dictionary of human motion manifold elements will be studied, which can integrate multifaceted information for analysis of human behavior. Abnormal human behavior recognition system based on the video surveillance system will be implemented, which can be used as the application and verification platform for the human behavior modeling method based on the library of human motion manifold elements. Project research will further extend the theory and application of human behavior recognition.

英文关键词： human behavior recognition;action recognition;Manifold Learning;Ensemble Learning