中文摘要：

自然环境下的日常动作识别有着广泛的应用前景和重要的研究价值.不同于以往在结构化和孤立条件下进行的动作识别,自然环境下的日常动作是连续的,视角多变并常发生遮挡.本文提出了分布式视觉系统下日常动作的在线识别方法.时间轴上的滑动窗口每个时刻取一段视频帧,采用基于＂包容形状＂的视角无关的体态表示方法提取体态特征向量,并用隐马尔科夫模型进行识别.动作类型的搜索空间由环境知识推理得到.遮挡检测和部分遮挡下的体态表示也在文中进行了讨论.实验表明本文提出的日常动作的在线识别方法能够克服日常场景给动作识别带来的困难,结果证实了方法的有效性.

英文摘要：

Recognition of actions in daily living is challenging because： （ 1 ） actions are continuous; （2） human location is changeful; （3） human body is partially occluded sometimes. This paper proposes a multi-view framework for on-line recognition of actions in daily living. Action representation based on ＂Envelope shape＂ enables view-invariant action recognition. A sliding window concatenates the feature vectors for action representation into a stream as the input to a bank of HMMs. A maximum likelihood based classifier detects action. The HMMs are chosen by an ontology based enviroument knowledge model. Besides, occlusion detection and action representation in occlusion are also discussed. Implementations demonstrate the efficacy of our approach.