【作者】 张军；

【导师】 刘志镜；

【作者基本信息】 西安电子科技大学 ， 计算机应用技术， 2009， 博士

【摘要】 智能视频监控是利用计算机技术实现自动视频分析的技术。作为一种有效安防的手段,智能视频监控系统正越来越受到人们的青睐。对于视频序列中行人异常行为的分析识别是智能视频监控领域中日益受到重视的一个研究方向。基于异常行为分析的智能视频监控系统不仅能忽略大量监控系统中对安防无用的信息,从而高效地完成保障安全的任务,还能节省大量的人力物力,给社会带来很大的经济收益。并且能够实现即时报警,解决了传统监控系统的事后性。本文从理论和实际应用的角度,对以视频为输入的异常行为识别进行了一些新的探索,以下所提出的算法均在实践中进行了应用。本文的主要贡献如下:1)提出了基于模板匹配的最小标准方差行为分类算法,解决了监控中不同人的类似动作但不完全相同的情况,有效解决了非固定速度的行为分类问题,能比较好地执行行为的分类任务。围绕着人的动作是时间序列上姿态的集合这样的事实,利用人的动作行为的周期特性,先将图像序列转换为一组静态形状模式,然后在识别过程中和预先存储的行为标本相比较。在建立标准库的过程中,提出了在最不相似的标准模板的生成算法,利用Procrustes中值形状距离作为图象的相似度标准,找出在一个视频测试周期内与标准库内动作标准方差最小的行为,实验证明最小标准方差算法解决了非连续性识别的要求,克服了模板匹配对运动时间间隔变化敏感的问题;提出了基于模板匹配的平均值与最小方差加权的行为分类算法,利用Zernike速度矩距离作为图象的相似度标准,找出在一个视频测试周期内与标准库内标准动作平均值与最小方差加权最小值,从根本上解决了在同一段时间内不同动作变速动作的分类,解决了某个图象集合中标准方差最小,但可能两个图象集最不相似的问题,增加了动作分类的准确率。2)提出了基于模糊联想记忆网络的行为分类方法,解决了动作姿态之间联系的问题,使得计算机识别过程向人的思维靠拢。把每个静态姿势作为一个状态,这些状态之间通过某种概率联系起来。任何运动序列可以看作为这些静态姿势的不同状态之间的一次遍历过程,在这些遍历期间计算联合概率,其最大值被选择作为分类行为的标准。人的运动轮廓特性被用作学习和识别的低级特征;学习是利用HMMs来为每个类别产生行为矩阵,作为模糊联想网络的知识,通过对实时视频流的认知,利用神经网络最终判断行为的种类,在不同情况下模糊联想网络的知识还可以随时学习更新。实验结果表明,算法得到了较好的识别结果,并具有一定的抗噪性。3)提出了基于模糊理论的判断异常行为识别方法,解决了判断前需要先定义异常行为的问题,实现了对异常行为的直接判断。首先为模型化人体结构设计了简化的人体关节模型图;其次根据行人躯干和四肢轮廓角度的变化,设计了用于模糊化的函数式;再次提出了利用躯干和四肢的模糊隶属度通过计算来得到整个人异常度的一种基于模糊理论异常行为判别的算法;最后在系统实现中,提出了利用质心轨迹和模糊判别的联合方法来甄别行人是否异常的方法,模糊判别可以实现在视频监控范围内对行人行为的主动分析,从而能够对行人异常的动作做出识别并进行报警处理。通过实验证明该方法具有较高的识别率。4)提出了可变场所的异常行为识别方法,解决了在监控应用中不同场景具有不同的异常判断规则的问题,使异常判断算法能应用在多种场所。采用双层词包模型判断在不同场景中行为是否为异常,把视频信息放在第一层包中,把场景动作文本词放在第二层包中。视频由一系列时空兴趣点组成的时空词典表示,动作性质由在指定场景下的动作文本词集合来确定。使用潜在语义分析概率模型(pLSA)不但能自动学习时空词的概率分布,找到与之对应的动作类别,也能在监督情况下学习在规定场景下运动文本词概率分布并区分出对应异常或正常行动结果。经过训练学习后,该算法可以识别新视频在相应场景下行为的异常或正常。5)提出一种能自动选择人体最大特性区域的覆盖比算法,解决了目前行人跟踪方法中,跟踪区域需要人工事先设定的问题。人体运动属于非刚性物体的时变和空变问题,在运动中的轮廓也是不断变化的。利用覆盖比算法能自动找到行人在初始时刻所需跟踪的最佳区域,然后选取特性区域内的加权颜色直方图作为跟踪特征,利用Bhattacharyya距离描述颜色模型的相似性,作为粒子权值的有力依据,最后在粒子滤波理论框架下实现自动地对行人进行实时跟踪。更多还原

【Abstract】 Intelligent Video Surveillance (IVS) is a kind of technology to achieve automatic video analysis with computer techniques. As an effective means of defense and security, IVS systems are being more and more popular. The analysis and recognition of pedestrian’s abnormal behavior in the video sequence, a research objective of IVS, has gradually drawn the attention in the field of IVS. The IVS system based on the analysis of abnormal behavior can not only ignore a large number of useless information, which guarantees the high efficiency in the security protection, but also save a lot of human and material resources, which brings great economic benefits to the whole society. In addition, it is also able to achieve real-time alarming to eliminate the lags in tradition monitoring systems. This thesis, in both the theoretical and the practical perspective, probes into abnormal gait recognition with the videos as input. The following proposed algorithm and methods are carried out in practice application. The main contributions of this thesis are summarized as follows:1) The classification method of the minimal standard deviation based on template is presented, which can solve the problem that different people have their special gait, and thus implement the classification actions better. Around the fact that people’s actions are a posture set in time sequence, the image sequences are firstly converted to a set of static shape mode in use of the periodicity of the human movement. And then in the recognition process the mode is compared with behavior of pre-stored samples. In the establishment of standard library stage, the minimal similar degree template generation algorithms is presented, which can solve the template representation problem. Take the Procrustes medium value shape distance as the similarity standard of the image, we can find out the behavior which has the minimal standard deviation from the standard database in one single test cycle. Experiments prove that the minimal standard deviation algorithm can meet the non-continuous identification requirements, and overcome sensitivity of the time interval changes in template matching. The weighted average of the maximal mean value and the minimal standard deviation algorithm is proposed, which take the Zernike speed matrix distance as the similarity standard of the image, can solve the uncertain speed motion and increase the precision of movement classification fundamentally.2) A Fuzzy Associative Memory (FAM) networks using behavior classification is proposed, which solve the problem between their movements and appearance, so that the computer recognition process can get closer to the people’s thought processes. Each static posture is treated as a state, these states are linked through some kind of probability. Any movement sequences can be treated as a traversal process of a static posture between the different states. The joint probability is calculated during the traversal, and the maximum value is chosen as the behavior classification standards. Pedestrian contour is used as a feature with study and identify in low-level. Behavior matrix for each category is generated with HMMs study. The motion classification is deduced by the knowledge of FAM network. In different circumstances, FAM network knowledge can be updated at any time through learning. This thesis uses four layers fuzzy neural network model, which is a system with multiple input and single output. It has nine input unit, which are eight standard deviations of each action and one centroid. Each input is a membership function. The first layer of the system is the input layer. The second layer is the membership function, whose effect is to turn the input into the membership degree. In the third layer, each node represents a rule that comes from the study algorithm. The relationship between present nodes and previous nodes relies on the rules; the node function relies on the application of the rules. The fourth layer is the output layer. In the system with multiple input and single output, the beginning weight is the membership degree of the rules. Then behavior breed has been obtained through the iteration algorithm. Experimental results show that the algorithm has given a good recognition results, and has a certain degree of noise immunity.3) Abnormal behavior of pedestrian detection based on Fuzzy theory is proposed, which solve the problem that the definition of abnormal behavior should be defined before determine, thus reach the direct judgments about abnormal behavior. Subject to certain scenes, scholars in and abroad have proposed methods based on statistical techniques, physical parameters, time-pace movement and model separately. The method of statistical techniques is robust and has fast calculation speed. The method using physical parameters is understandable and observation angle independent, while it depends on the parameter of the recovered scene. The time-space movement method can reveal the character of the time and space, but can easily be disturbed by the noise. The model method has the problem that we can hardly get the precise model from the video. It also calls for massive amount of processing. To detect and tract the particular moving targets in specific environment, our method is unrelated with the scenes. Firstly a simplified human joints model has been established to model the human body. Then a fuzzification function is designed with the variety of body’s trunk and limbs contour angles. Thirdly an abnormal behavior discrimination algorithm based on fuzzy theory is proposed, which applies fuzzy membership of the pedestrian’s trunk and limbs to get the overall degree of the anomaly. Finally in the reality of the system, a combined method of center of mass and fuzzy discriminant is presented. Fuzzy discriminant can detect irregularities and implements initiative analysis to body behavior in the visual surveillance. Therefore, abnormal behaviors can be recognized and alarmed. The results show that the new algorithm has a high recognition rate.4) A method of abnormal action recognition in variable scenarios is proposed, which eliminate the ambiguity that one single action of the same person can lead to different comprehension under different circumstances. In the monitoring application, different scenarios have different exceptions to determine the rules of the algorithm and can be applied to determine abnormalities in a variety of places. There are different understanding results in different scenarios even if the same person’s action in visual analysis. In order to determine whether the behavior is abnormal in different scenarios, a double-layer bag-of-words model is proposed to solve the problem in our surveillance system. The video information is processed in the first layer of bag-of-words, and the information of scenario-action text words is included in the second one. A video sequence is represented as a collection of spatial-temporal codebook by extracting space-time interest points. The behavior characteristic is represented as a collection of behavior text words in special scenarios. Probabilistic Latent Semantic Analysis (pLSA) model is adopted to automatically learn the probability distributions of spatial-temporal words and the topics correspond to human action categories. PLSA can also learn the probability distributions of the motion text words in a scenario with supervisor and the topics correspond to anomalous or normal actions. The algorithm can categorize the human anomalous or normal action contained in the special occasion to a novel video sequence after being trained.5) A method of automatically selecting characteristics for pedestrian tracking is proposed. The paper presents an algorithm of cover ration which can select the largest feature region of human body automatically to improve the way that the tracking area is pre-designated in current pedestrian tracking methods. Since human motion belongs to the time-varying and space-varying problem of a non-rigid object, the outline of a pedestrian in motion is changing constantly. The algorithm of cover ration can find the best area of the pedestrian which can be tracked at any time automatically. Then it can select a weighted color histogram within the feature region as tracking features and take the similarity of color model described by Bhattacharyya distance as a strong evidence of particle weight. Finally, in the framework of particle filter theory, the real-time tracking of pedestrians can be achieved automatically.更多还原