【作者】 谢非；

【导师】 徐贵力；

【作者基本信息】 南京航空航天大学 ， 检测技术与自动化装置， 2009， 硕士

【摘要】 智能视觉监控技术是计算机视觉领域一个新兴的应用方向和备受关注的前沿课题,监控的目的常常是对监控场景中的异常情况或监控对象的异常行为进行检测、分析,因此,对人的多种姿态进行识别有着十分重要的意义。本文针对户外智能监控中车辆、人体直立、哈腰、匍匐三种姿态的识别方法进行了研究。首先,为了便于车辆及三种人体姿态识别的研究,建立了样本库,每种目标均采用了不同大小、多种方位的实物进行拍摄,每种样本的数量为200幅。为了进行多种目标识别方法的研究,提取了多种特征参数,如宽高比、惯性主轴角、矩形度、形状复杂度、离心率等,并研究了多种细化算法及轮廓提取算法,细化算法包括OPTA细化算法、Zhang细化算法、Rosenfeld细化算法、横井法细化算法,在对四种细化算法进行比较后发现横井法细化方法速度较快,细化效果较好,但存在细化不太彻底及细化不平滑的问题,因此,对其进行改进,改进后的算法细化效果更加彻底、平滑,利用细化后的目标及原目标二值图设计了双腿搜素算法及头部搜素算法,从而为后续决策树式识别方法的研究做好了准备。其次,为了寻找到一种较好的识别方法,对一般常用的识别方法进行了研究,包括多种不变矩方法、RBF神经网络分类法及支持向量机分类法,实验表明多种不变矩方法不适于多种人体姿态的识别,而RBF神经网络及支持向量机方法的分类性能对运动目标特征量的依赖性较大,在初次尝试提取三种特征量进行分类后,发现正面哈腰人体与直立人体不能较好的区分。因此,提出了基于投影直方图的特征量提取方法,利用此特征量进行训练、分类,可得到较好的识别效果,平均识别率达到96%。此外,由于神经网络及支持向量机的内部结构较为复杂,设计实现时步骤较为繁琐,而且需要大量样本进行训练。因此,提出了一种结构简单、便于理解的识别方法研究,基于决策树式的识别研究。最后,由于决策树具有把复杂分类问题分解为多个简单分类问题的特点,同时可以把人的经验知识加入其中,采取小样本就可设计实现,所以又进行了基于决策树的人体及其姿态的识别研究。识别原理为:首先利用矩形度与形状复杂度变化率将车辆与人体区分出来,然后使用宽高比粗略估计出人体目标,进一步搜素人体双腿或头部从而确定人体,最后利用头部与腿部位置的匹配判定人体姿态,系统通过多帧综合判别的结果最终确定了目标种类及人体姿态。考虑到监控系统中多目标同时出现的几率较大,因此在单目标识别的基础上,提出了多运动目标的识别方法,利用对各目标轮廓跟踪的方法将其标记,单独提取后进行识别分析,达到多目标中单目标提取、识别的目的。利用样本库对决策树式识别方法进行测试,实验结果表明,该方法具有算法简单,识别率高,平均为96%,识别速度快,每帧识别时间平均为0.4322s,以及多方位人体姿态识别的特点。为智能监控系统中人体行为理解的任务做了基础研究工作。更多还原

【Abstract】 Intelligent surveillance technology is a new application and hot research subject in computer vision area. The aim of surveillance system is detecting moving object and analysising abnormal situation or abnormal object in surveillance scene, so the recognition of various human posture is very important. This paper explores the means of recognition of cars and three human body postures, including standing body, stooping body and creeping body, in outdoor Intelligent surveillance system.First, in order to make research on the recognition of cars and human body postures conveniently, we build a sample library for various size and orientation moving objects, every kind object has the number of 200 samples. Many character parameters have been extracted, such as the ratio of width and height of object, inertial spindle angular, rectangle degree, shape complexity degree, eccentricity, etc for the research of many recognition methods. And many thinning algorithms and two silhouette tracking algorithms have been introduced for the design of human head and legs searching algorithms, including OPTA thinning algorithm, Zhang’s thinning algorithm, Rosenfeld’s thinning algorithm, Hengjing thinning algorithm. After comparing the four thinning algorithms, we find the hengjing thinning algorithm has a good thinning performance and needs less execution time, but it also has a problem that it doesn’t have a complete and smoothing thinning effect. Having improved the hengjing thinning algorithm, the object can be thinned thoroughly and smoothly. The human’s head and legs searching algorithm prepares for recognition method research based on decision tree later.Secondly, in order to develop a better recognition method, we make research on the common pattern recognition methods, including a variety of invariant moments method, RBF neural network classification method, SVM classification method. Experimental results demonstrate the first method is not very useful in presence of classification and the performance of the last two methods are obviously strictly related to the selected features. Having selected three features to test, we find a problem which is the misclassification between the front stooping human body and creeping human body posture. Therefore, we propose a method of choosing features based on the histograms of horizontal and vertical projections. The selected features have a large discriminatory capability and a good recognition ratio, which is up to 96%. Moreover, considering the complicated structure of the neural network and SVM, it is hard to design and needs lots of samples to train. So we present an approach with a simple structure, the decision tree system.Finally, for the decision tree has advantages of making complex classification problem solved simply and adding the experiences of human to the decision, so the decision tree recognition system can be designed using a small number of samples. The recognition principle is distinguishing the car and human by rectangular degree and shape complexity degree firstly, and then recognizing the human object roughly using the ratio of width and height of the object, at last confirming the human body through the head and legs detecting algorithms and judging the human posture by location of human’s head and legs. The system using many frame recognition results as the system recognition results, which confirms the object class and the human posture. Considering the high ratio of multiple moving objects appearing, we present a multiple objects recognition methods based on the single object recognition method. Marking each object through tracking its silhouette, and recognizes it after extracting the single object. At last we realize the purpose of extracting single object and recognizing it in multiple objects appearing situation.Testing the recognition system based on decision tree, experimental results show that, this approach has a simple algorithm, a good recognition ratio, which ups to 96%, a fast recognition speed, which is 0.4322s, and a advantage of many orientation of human body posture recognition. This paper provides a fundamental research for human body activity understanding task in intelligent surveillance system.更多还原