【作者】 黄彬；

【导师】 田国会；

【作者基本信息】 山东大学 ， 控制理论与控制工程， 2010， 博士

【摘要】 目前许多国家已进入老龄化社会,对独居老人的监护和陪伴已成为突出的社会问题,各国把智能服务机器人的开发作为应对这一问题的重要策略之一。正确识别和理解人的行为是服务机器人给人提供相应服务的前提。目前对基于视觉的行为识别和理解的研究尚处于初级阶段,人体行为模式的多样性和家庭环境的复杂性使之成为一个极富挑战性的课题。就目前机器人的发展水平而言,仅仅依靠机器人自身有限的感知和计算能力难以完成“在正确识别和理解人的行为基础上提供恰当服务”这样复杂的任务,必须借助其他手段。为此,本文提出了以基于视觉的方法为基础、结合智能空间技术实现对人行为的识别和理解的方法。对智能空间中人的行为的正确识别和理解使机器人对独居老人的监护和陪伴成为可能。因此,对智能空间及智能空间中行为识别和理解的研究具有重要的理论和实际意义。本文对智能空间中行为识别和理解所包含的五个主要方面进行了研究与探讨：“面向理解和服务的智能空间”(Intelligent Space for Understanding and Service,简称ISUS)的设计与搭建,运动目标的检测与跟踪,日常动作识别,异常动作识别,行为认知和意图识别。首先,研究了ISUS的搭建。ISUS既是人活动的物理空间,又是能够感知人的活动、给人提供服务的智能空间。在行为识别和理解的研究过程中,需要用到ISUS中的视觉和其他相关信息。因此,ISUS是后续研究的基础；然后,研究了以家庭环境中独居老人的监护为背景的目标检测和跟踪,实现了家庭环境中对目标的准确检测和实时跟踪,此为目标动作识别中特征提取的前提；在正确检测出目标并对目标进行实时跟踪的基础上,对常见日常动作和异常动作的识别进行了研究。日常动作的正确识别是行为认知和意图识别的基础,而通过对异常动作的有效识别可判断被监护者是否有异常状况发生；最后,研究了智能空间中目标的行为认知和意图识别,实现了对正常行为的认知和意图识别以及对常见反常行为的识别。在此基础上,智能空间可根据识别结果直接或者间接给人提供相应的服务。主要研究内容和结果概括如下：(1)针对iSpace (Intelligent space)服务效率低的问题,提出了ISUS的概念和执行分布的思想,有效提高了服务效率,降低了机器人的负担；从智能空间标准化的角度进行研究,设计了ISUS的分层体系结构,从而使整个系统更加紧凑清晰,便于ISUS的搭建和任务分配；在对ISUS的组成单元进行研究、设计的基础上,搭建了智能空间,然后针对智能空间如何有效完成“理解和服务”这一核心任务进行了任务分解和规划,最后介绍了ISUS所能实现的功能并剖析了ISUS中的信息流结构。(2)针对智能空间的环境特点,改进了高斯混合模型(GMM)运动检测算法,提高了模型及时反映背景变化的能力,从而改善了目标分割的效果；对均值漂移算法(Mean Shift算法)进行了改进,提出了适合本环境的模板和目标尺度更新策略,并结合卡尔曼滤波提高了跟踪的鲁棒性；采用智能空间中的无线传感网络技术和视觉技术相结合的目标跟踪方法,解决了跟踪过程中目标的严重遮挡或目标暂时离开摄像机视野范围后的重新获取以及多摄像机目标跟踪中目标匹配、交接困难的问题,实现了多摄像机对目标的接力跟踪。(3)人体动作识别是行为认知和意图识别的基础,基于视觉的动作识别的实质是有效特征的选择问题,好的特征不仅能够有利于行为识别而且能够减少识别的计算量。针对利用目前常用特征难以提高动作识别准确性的问题,本文提出了利用人体动作序列的轮廓特征识别人的日常动作的方法,有效提高了动作识别的准确率。但是,这种方法对目标分割的要求很高,易受目标分割结果的影响。为此,提出了直接利用人体动作序列的侧影特征作为动作识别特征的方法,提高了识别的准确性和可靠性。(4)对常见异常动作的识别进行了研究,主要研究了跌倒、异常步态和异常行走状态的识别。在跌倒识别方面,针对已有方法难于区分跌倒和正常躺下的问题,提出了一种新的跌倒识别算法,在识别过程中利用了跌倒的静、动态特性,同时考虑了人在智能空间中的位置、身体状况等信息,提高了跌倒识别的准确性；与目前常用步态特征实现身份识别不同,本文通过对异常步态的识别来推断人是否有异常状况发生。为此,根据人的步态特点,定义了能实时描述步态变化的参数,据此实现了对异常步态的识别；最后通过对行走轨迹的判断实现了对异常行走状态的识别。(5)行为认知和意图识别是行为识别和理解的高层处理阶段。首先对人在智能空间中的定位进行了研究,针对基于身高模型的定位不能解决当人的姿态变化时会产生定位错误的问题,提出了身高模型和姿态识别相结合的定位方式,消除了人体姿态变化对定位准确性的影响。然后,通过定义的位置置信度系数实现了多摄像机定位结果的融合,进一步提高了定位精度；在实现对人定位的基础上,研究了对正常行为的认知和对反常行为的识别。在对正常行为的认知和意图识别中,提出了智能空间关键点的概念,对智能空间中的关键点和区域进行了定义。在此基础上提出了地点驱动的行为认知算法,把关键点与其属性相联系,实现了对人意图的识别；在反常行为识别方面,提出了通过分析人在关键点之间的活动路线识别人是否有反常状况发生的方法；提出了关键点持续时间直方图的概念,并利用持续时间直方图实现了对反常习惯行为的识别,并通过定义的习惯变化评价指标实现了对行为反常程度的度量。更多还原

【Abstract】 Nowadays, the medical care and accompaniment of solitary senior citizens has become a special social problem due to the aging society tendency in many countries. One accredited strategy to deal with this crisis is to develop intelligent service robots. Particularly, human behavior recognition and understanding are two essential prerequisites for service robots to function correctly. At present, the research of behavior recognition and understanding based computer vision is still in its preliminary stage. It is an extremely challenging task for the robots to adapt to the variety of human behavior patterns and the complexity of home environment. In other words, it is very difficult to accomplish such complicated mission as providing correct service based on the recognition and understanding of human behavior as far as the currently robot development level is concerned. Therefore, in order to solve this kind of problems, a vision-based method is presented in this thesis that can achieve human behavior recognition and understanding combined with the intelligent space technology. This method makes it possible to accompany and take care of solitary elderly based on the intelligent service robots and the intelligent space technology. It can be shown that the study of the intelligent space and behavior recognition and understanding in intelligent space have important theoretical and practical meanings.In this thesis, five essential aspects concerning behavior recognition and understanding in intelligent space are studied which include the research and construction of ISUS (Intelligent Space for Understanding and Service), detection and tracking of moving object, typical action recognition, abnormal action recognition, and behavior recognition and understanding. Firstly, the construction of ISUS is studied. ISUS is not only a physical space of human activities, but also an intelligent space which can perceive human’s behavior and provide assistance to human in it. The visual and other relevant information of ISUS is required in the course of study on the behavior recognition and understanding. Therefore, ISUS is the foundation of the following research. Moreover, object detection and tracking in the background of the old solitary people in home environment is studied. The accurate target detection and real-time tracking are two prerequisites for feature extraction in the action recognition which will be shown in the following. Based on the correct detection and tracking of objects, the recognition of common daily actions and abnormal actions can be achieved. The correct recognition of the daily actions and abnormal movements are foundations of the behavior perception and intention recognition. Lastly, behavior perception and intention recognition in ISUS is researched, and the intention recognition of the normal behavior and the abnormal behavior recognition are achieved. Based on the above five aspects, ISUS can provide the corresponding service to human directly or indirectly according to the recognition result. The main research contents and results are shown as follows:(1) In view of the inefficiency of iSpace, the concept of ISUS and the idea of execution distribution are discussed, which are used to effectively improve the service efficiency and reduce the burden of the robot. The layered architecture of ISUS is designed from the perspective of intelligent space standardization, which makes the system more compact, clear and convenient to be built. Meanwhile, this method is convenient for task allocation. Moreover, the constructions of ISUS, task planning and information flows composition in ISUS are studied based on the above researches.(2) The GMM algorithm is improved according to the environment characteristic of ISUS. It improves the capability of models to timely reflect the background changes. Therefore the effect of target segmentation is increased, the mean shift algorithm is improved and a template and target scale updating strategies for the ISUS environment is proposed. The robustness of object tracking is raised combined with a Kalman filter. In this paper, object locationing and tracking technology combined wireless sensor networks and computer vision is adopted. Many problems, such as the heavy occlusion and re-capturing of objects during object tracking, the difficulty of target hand off and object matching in multi-camera object tracking, are solved. Object relay tracking in multi-camera system is also realized with this method. (3) Human action recognition is the foundation of behavior perception and intention recognition, and the feature selection is a key to it. Good features can not only be conducive to recognition, but also reduce the amount of calculations. In view of the problem that it is difficult to improve the recognition accuracy using the common features, this thesis presents a methodology for typical action recognition using the contours of human posture, which effectively improves the action recognition accuracy. In order to overcome the shortcomings of human body contour apt to be effected by segmentation results, the silhouette of human action is used as recognition feathers after dimension reduction which improves recognition reliability.(4) The abnormal action recognition in home environment is studied which includes the falling, abnormal gait and abnormal walking status. Especially, as it is difficult to distinguish falling from lying down using existing methods, a new algorithm is proposed in falling recognition, which increases the accuracy of falling recognition using the static and dynamic features and human location in ISUS and physical condition. Moreover, different from the gait-based human identification, the purpose of abnormal gait recognition in this thesis is to infer whether an abnormal situation has occurred. Therefore, gait parameters reflecting stride changes are defined by the human gait characteristics and the abnormal gait recognition that can be achieved by analyzing these parameters. Lastly, the abnormal walking state recognition is achieved by judging human walking trajectory.(5) Behavior perception and intention recognition belong to the high-level processing stage. In this thesis, the human locationing in ISUS is researched first. A new algorithm of human locationing in ISUS is proposed using height models combined with posture recognition, which reduces the effects of body posture changes to locationing accuracy. The locationing accuracy is further improved using the combination of multi-camera locationing results by the defined position confidence coefficient. The normal behavior perception and abnormal behavior recognition are researched based on human location. In normal behavior perception and intention recognition, the concept of key points in ISUS is discussed, and key points and area of ISUS are defined. Then a new behavior perception algorithm based on place-driven is proposed, which achieves human’s intention inference through interrelating key points and its attribute. In abnormal behavior recognition, a key point based method of abnormal event recognition using activity route is introduced. Then the concept of key point’s duration histogram is put forward and used to recognize abnormal behaviors. Finally, the abnormal degree of a behavior is obtained by the defined evaluation index of habit changes.更多还原