【作者】 任海兵；

【导师】 徐光祐；

【作者基本信息】 清华大学 ， 计算机应用技术， 2003， 博士

【摘要】 本论文研究的内容是非特定人自然的人体动作识别。“非特定人”的含义是：无需对每个动作主体单独建立特定的动作模型；“自然的人体动作”指人们日常生活中的动作，系统无需对动作主体进行任何训练。在智能的人机交互、智能监控、基于内容的检索和基于人体模型的超低比特流的压缩算法等方面，非特定人自然的人体动作识别有着广阔的应用前景。但是，因为人体是复杂的多关节变形体，精确地获取人体运动参数非常困难，而且更为主要的是非特定人自然的人体动作参数离散度很大，利用传统的识别算法很难得到令人满意的结果。所以，非特定人自然的人体动作识别是一个极富挑战性的研究课题。针对传统算法的不足，本文提出了以基于基元特征的动态贝叶斯网络为核心的非特定人自然人体动作识别策略，较好地解决了运动特征获取困难、参数离散度大的难题，显著地提高了系统识别率和鲁棒性。研究工作主要包括以下四个方面：1. 人体运动的基元特征表示法为表示人体运动的本质属性，本文提出了基元特征的概念。由高层知识归纳推理得到的基元特征，突出地描述了与人体动作相关的上下文信息、人体运动过程和姿态的显著特征。在非特定人自然的人体动作识别中，基元特征能够大大降低参数的离散度，提高系统抗噪声能力、鲁棒性和识别率。由于基元特征参数的维数低，因此，它与基于状态的各种识别方法相结合有利于实现更复杂识别方法。2. 基于动态贝叶斯网络的多信息融合识别策略本文首次引入动态贝叶斯网络来融合多种信息推理，进行自然的人体动作识别。在动态过程识别中，动态贝叶斯网络可以根据具体问题设计具体的动态贝叶斯网络结构。这样的网络结构不仅可以将各种信息融合推理，使得多个弱信息达到强信息的作用，而且可以根据信息的来源、置信度、重要性等设置不同的权重，使信息配置更加优化合理，有效地解决复杂概率推理问题。本文将基元特征与动态贝叶斯网络相结合，得到了兼具两者优点的基于基元特征的动态贝叶斯网络，从而在非特定人自然的人体动作识别中大大提高了<WP=5>系统的识别率和鲁棒性。3. 背景自动补偿和背景映射相结合的人体区域分割算法从图像中精确分割人体区域必须消除摄像机光圈自动调节对背景成像的影响、去除人体阴影区域。针对光圈自动调节对背景成像的影响，本文提出了背景自动补偿的前景分割算法，通过背景图像自动补偿来消除光圈影响。针对人体阴影区域，本文提出了基于背景映射关系的快速分割方法，利用两个视图背景图像之间的映射关系快速去除人体阴影区域。这两种方法的结合不仅可以精确地分割出人体区域，而且减少了计算量，提高了鲁棒性。4. 基于全局信息的特征检测方法在特征抽取方面，本文提出一种基于全局信息的特征检测方法。该算法克服了传统算法仅能抽取双手运动信息的缺点，利用人体剪影全局信息来获取人体平移、人体旋转、肘关节夹角、双手速度等多方面的运动信息，因而能够更加详细地描述非特定人自然的人体运动，进而提高识别算法的分辨能力。本文通过大量的实验比较了隐马尔可夫模型、耦合隐马尔可夫模型、基于基元特征的耦合隐马尔可夫模型和基于基元特征的动态贝叶斯网络这四种模型的识别效果。实验结果表明：基元特征可以大大降低非特定人自然人体动作特征参数的离散度；而动态贝叶斯网络能够有效地融合多种信息进行推理，从而显著提高自然人体动作的识别率。更多还原

【Abstract】 This thesis focuses on subject-independent natural action recognition. Here, "subject-independent" means that it is unnecessary to build specific action models for each subject; "natural action" refers to daily actions with no training needed for each subject. It has broad applications in intelligent human-computer interaction, smart surveillance, content-based retrieval, model-based very low bit-rate video compression, etc. However, as human body is a complex articulated object, it is difficult to obtain the precise motion parameters. More importantly, with large scatter in the motion parameters, it is nearly impossible to get satisfying results with conventional recognition algorithms. Therefore, subject-independent natural action recognition is a very challenging research topicTo overcome the demerits of conventional algorithms, subject-independent natural action recognition algorithm with Primitive-based Dynamic Bayesian Network as its core is proposed in this thesis, which can obtain more precise motion features and reduce scatter of the motion parameters, thus improving considerably the recognition rate and robustness. The research focuses on the following four points:1. Primitive-based Human Motion RepresentationThe concept of primitives is proposed in this thesis to represent the essential features of human motion. Inferred by high-level knowledge, Primitives are distinctive features which give a best representation of the system state and describe the context information related with human actions and the motion information representing human action state as well as pose. Primitives can reduce the scatter of the motion parameters greatly and improve the subject-independent action recognition rate and robustness.As the parametric representation of primitives is low dimensional, the integration of primitives and state-based recognition algorithms helps realize some more complex recognition models. 2. DBN-based Mulit-information Fusion for RecognitionDynamic Bayesian Networks are first introduced in this thesis for natrual action<WP=7>recognition. In the recognition of dynamic system, different structures of Dynamic Bayesian Network are designed for different problems, in which many kinds of weak information are fused to be strong information. And in accordance with the source, confidence and importance, each kind of information is given a weight, which is very useful for multi-information optimization and inference.In this thesis, primitives and Dynamic Bayesian Networks are combined to be Primitive-based Dynamic Bayesian Networks, which have the both merits, thus improving greatly the recognition rate and robustness of the system.3. Integration of Background Auto Compensation and Background Mapping for Human Silhouette SegmentationTo segment precisely human silhouette, it is imperative to remove the aperture auto-adjusting effect and delete the human shadow area. For the aperture auto-adjusting, background auto compensation approach is proposed in this thesis to remove the effect by auto-compensating the background image. To remove the human shadow area, a fast segmentation approach based on stereo background mapping is given. In this thesis, the integration of the two approaches not only segments the human silhouette precisely, but also reduces the computation and improves the robustness. 4. Global Information-based Feature DetectionOn feature detection, an algorithm based on global information is presented in this thesis. Compared with the conventional algorithms which can only get the motion information of two hands, this algorithm can obtain multi-information through the global information of human silhouettes, including human body translation information, rotation information, elbow angles and the two hand trajectories. Therefore, it gives a more detailed description of the subject-independent natural actions, thus improving the recognition ability.In the experiments, the recognition results of four models are compared, namely, HMM, CHMM, PCHMM and PDBN. The resul更多还原