【作者】 高琳；

【导师】 王世刚；

【作者基本信息】 吉林大学 ， 信号与信息处理， 2013， 硕士

【摘要】 运动人体的行为识别分析是智能视频监控的研究热点之一。它是计算机对视频内容的自动分析，作用是使机器有人一样的主观判断能力。其工作是从视频图像序列中检测提取运动对象，对对象进行跟踪，进而分析或描述运动对象的行为类别。本文主要工作是研究基于视频处理技术的行为识别，研究了三方面问题：运动对象检测与提取、运动对象跟踪、运动对象行为分析。在运动对象检测与提取方面，采用基于高阶统计法检测运动变化信息时，提出了一种阈值背景窗口自适应选取的方法，并通过基于二次帧间差分法的累积帧间差分法提取对象。对于运动对象跟踪，提出采用基于特征匹配和卡尔曼滤波相结合的对象跟踪算法。对于运动对象行为识别，提出采用了基于支持向量机的运动对象行为识别，并提出了自己的特征描述方法。运动对象检测和提取过程：首先分别计算前后两次累积帧间差分，然后分别进行运动变化检测，滤除噪声，再对前后两次处理后的图像取交集，提取出运动对象。在基于高阶统计法检测运动变化信息时，提出了一种阈值背景窗口自适应选取的方法，此方法能够使系统自适应的变换阈值。运动对象跟踪过程：对检测提取到的运动对象，提取其位置、形状和亮度特征，组成各个对象的特征向量，并开始标号跟踪。对当前的视频帧进行运动对象检测提取，并与前面一帧进行连续两帧之间的对象的特征匹配。对未能成功匹配的对象，讨论可能出现的情况，如遮挡、消失和新对象出现等。首先讨论研究遮挡情况，若对象之间判断为发生遮挡，则采用卡尔曼滤波算法预测对象在当前帧可能出现的位置，进行特征匹配；若前一帧有对象在当前帧没有匹配成功，同样采用卡尔曼滤波预测其在当前帧可能出现的位置区域，再对这连续两帧之间没有匹配成功的区域匹配；依然无法匹配成功的对象，对其标记为新对象。结束每一帧的跟踪之后，要更新运动对象特征信息，来进行后续帧的对象跟踪。运动对象行为识别的过程：对于跟踪的运动对象的特征，选择支持向量机作为行为类别的分类器，在训练分类器时，论文提出了基于时序的特征描述：选取运动对象在每一帧的位置特征，计算对象在帧与帧之间移动的距离，并研究这种移动的距离在时间，即连续帧上的统计特性，作为特征。通过特征样本数据库的建立，分析其可分类机制，设计分类器并训练分类器，以达到良好的分类效果。最后通过实验仿真结果，验证了本文行为分析系统的各分类器的准确率。更多还原

【Abstract】 Behavior recognition of human movement is one of the central issue of theintelligent video surveillance. It enables the computer automatically analyzes thevideo content and lets the machine have the same subjective judgment with human.The main job is to detect and extract the moving object from the video sequence, andto track the object. And then the computer can analysis or descript the category of themoving object.The main work of this paper is to study the behavior recognition based on videoprocessing technology. There are three aspects in this paper: moving object detectionand extraction、moving object tracking and moving objects behavior recognition.On the aspect of moving object detection and extraction, propose a method that is theadaptive selection in background window of a threshold when detect changes inmotion information based on higher order statistics method. This paper will apply theCumulative inter-frame difference method based on Secondary inter-frame differencemethod to extract object. On the aspect of moving object tracking, a method based onthe combination of feature matching and Kalman filter is proposed. On the aspect ofmoving object behavior recognition, put forward a method based on Support VectorMachine to recognize the behavior of moving object and propose my owncharacterization methods.The process of moving object detection and extraction: First of all, calculate thecumulative inter-frame difference before and after the current frame. Secondly, detectthe changes of movement respectively and filter out the noises. At last, take theintersection of the cumulative inter-frame difference before and after the currentframe to extract the moving object. When apply the method of detecting changes inmotion information based on higher order statistics, this paper proposes a method thatis the adaptive selection in background window of a threshold, it can enable thesystem select background window of a threshold adaptively.The process of moving object tracking: After the detection and extraction, we getthe moving objects. Extract the position, shape and brightness characteristics of theseobjects. Use these various characteristics compose the feature vector of each objectand label each object to start the tracking. Detect and extract the moving objects incurrent frame of the video, match the feature of the moving objects between currentframe and the frame in front of it. Since there maybe objects failed to match, wediscuss the possible condition, such as occlusion、disappearance and the appearance ofa new object. First we discuss the condition of occlusion. If the occurrence ofocclusion is judged between objects, use the Kalman filter algorithm to predict thepositions of these objects in the current frame, and match their features. And then ifthere are objects of the front frame failed to match in current frame, also use theKalman filter algorithm to predict the positions of these objects of the front frame inthe current frame and match their features. However, if there are still objects failed tomatch in current frame, it may be the appearance of a new object and label this kind of objects. After the end of the tracking in each frame, update the feature informationof the moving objects to carry out a subsequent frame and track the objects.The process of moving object behavior recognition: This paper proposes the SupportVector Machine as the class classification system according to characteristics of thetracking objects. In the training of the classifier, the paper presents a time-basedcharacteristic to describe the behavior of objects. Select the position feature of themoving objects in each frame and calculate the moving distance of the objectsbetween two consecutive frames. Study the statistical properties of the distance as thefeature. Through the establishment of database with the feature of the samples,analyze the classification mechanism and design the classification system to achieve agood classification effect. At last through the simulation results, verify theclassification accuracy of this article behavior recognition system.更多还原