基于人-物互信息模型的HOI行为异常侦测机制研究

项目名称： 基于人-物互信息模型的HOI行为异常侦测机制研究

项目编号： No.61201429

项目类型： 青年科学基金项目

立项/批准年度： 2013

项目学科： 电子学与信息系统

项目作者： 蒋敏

作者单位： 江南大学

项目金额： 24万元

中文摘要： 心理学研究证明人-物交互作用在人体行为识别中具有非常重要的作用，但鲜有研究者在该方向做深入研究。本课题以人体异常行为识别问题为研究对象，提出了一种新的基于人-物互信息的异常行为侦测方法。主要工作包括：1）深入分析人-物共生关系，综合考虑人、物外观特征、几何特征，共生特征，引入刚体约束及人体动力学约束，融合多元信息，基于分层随机场思想构建人-物交互行为自适应表达模型；2）基于主动学习机制，引入原子动作字典，研究人体关节空间位置先验边缘分布特性及空间、动力学等多元约束，基于信息质心强化模型，反馈调整模型参数，完成样本池优化模型构建及新样本的分类策略智能调整；3）模拟人的自然思维模式，研究不同样本间行为、动作、目标物体之间的"距离"以表征行为的异常度，构建一种全新的APO距离异常行为度量机制。本项目研究有助于丰富和发展人体行为识别与理解的理论研究体系，并为异常行为侦测研究提供了全新的研究方向。

中文关键词： 人体行为识别；人物互关系；物体识别；深度图；3D骨骼

英文摘要： Although psychology studies have demonstrated that Human-Object interactivity plays an important role in human behavior recognition, this research area hasn't been explored thoroughly. According to the human abnormal behavior detection problem, A new method based on human-object mutual information is proposed in this project. The project includes three main jobs: 1)By deeply analyzing the feathers of human, object, geometry and symbiosis relationship, taking the confinement of rigid body and kinetics of the human body into account, the human-object interbehavior self-adaptive model will be constructed based on hierarchical random field. 2)Based on the atom pose dictionary, the prior marginal distribution for joint location is studied with the confinement of space and kinetics. Considering active learning, sample pool optimal model is constructed with the parameters being adjusted through the feedback of information centroid reinforcement model which effectively shield the learning process from plenty of feedback noises. 3)To imitate the natural thinking pattern of human, the distance between actions, poses and objects (APO Distance) of different samples are discussed to represent the abnormal quantization of action. Upon the APO Distance, a new mechanism of HOI Abnormal Activity Detection is built. This researc

英文关键词： human action recognition；human-object interaction；object recognition；depth map；3D skeleton