运动阴影检测与目标识别方法研究

发布时间：2018-04-08 17:39

本文选题：智能视频监控　切入点：哈尔型特性局部二元模式　出处：《中国科学技术大学》2017年硕士论文

【摘要】：当今科技发展日新月异,社会经济水平稳步提升,人口的流动性日益增大,大量的流动人口给社会治安带来挑战,传统的安防监控系统依靠人力对场景视频进行分析处理,没有充分利用计算机技术以最大化视频监控的价值。近年来,计算机视觉、机器学习等技术不断发展,智能视频监控系统日趋成熟,在众多公共场所中得到了大量应用,相比于传统方法,它不仅能减轻监控人员的工作负担、节约成本,还能提高数据处理、异常检测的能力。结合当前先进的理论和技术,智能视频监控系统将在建设智慧城市、平安城市等方面发挥重要作用。由于监控场景的多样性和复杂性,目前智能视频监控还面临许多问题和挑战,本文主要针对运动阴影检测以及运动目标识别问题进行研究,主要工作如下:(1)分析了产生运动阴影的基本原理及其相关性质,介绍了不同类型的运动阴影检测算法。分析了经典的运动目标检测算法并对基于码本的方法进行了详细说明,对目标分类中常用的特征及分类器进行了介绍。(2)提出了一种基于哈尔型特性局部二元模式(Haar Local Binary Pattern,HLBP)特征的运动阴影检测算法。该算法分别提取运动区域及其背景区域的HLBP特征向量,不需要进行阈值选取、图像分块以及直方图统计,使用曼哈顿距离度量纹理差异,对差异图像进行阈值分割以检测运动阴影。结合颜色空间信息及纹理信息,提出了一种基于随机森林的运动阴影检测方法,该方法不需要对应用场景的光照、反射性质等进行假设,也避免了参数的设置。使用随机森林分类器对各个像素点进行二分类,判决是阴影还是前景目标像素点。该方法对各种室内外场景能取得较好的效果并且具有一定的泛化能力。(3)提出了一种基于HOG-HLBP特征的运动目标识别方法。利用方向梯度直方图(Histogram of Oriented Gradient,HOG)对目标外形轮廓以及HLBP对纹理的描述能力,将两者结合形成HOG-HLBP特征,使用支持向量机(Support Vector Machines,SVM)进行多类别判决,实验证明该方法能取得较好的分类效果。实现了一种基于检测的多目标跟踪方法,在匹配代价中加入表观信息,使其能够应对一些遮挡情况。结合运动检测、运动阴影消除以及多目标跟踪实现了监控视频目标分类。
[Abstract]:With the rapid development of science and technology, the steady improvement of social economic level, the increasing mobility of the population, a large number of floating population bring challenges to social security. The traditional security monitoring system relies on manpower to analyze and process the scene video.Computer technology is not fully utilized to maximize the value of video surveillance.In recent years, with the continuous development of computer vision, machine learning and other technologies, intelligent video surveillance system is becoming more and more mature, and has been widely used in many public places. Compared with traditional methods, it can not only reduce the workload of supervisors.Save cost and improve the ability of data processing and anomaly detection.Based on the advanced theory and technology, intelligent video surveillance system will play an important role in the construction of intelligent city and peaceful city.Because of the diversity and complexity of surveillance scene, intelligent video surveillance still faces many problems and challenges. This paper mainly focuses on moving shadow detection and moving target recognition.The main work is as follows: (1) the basic principle of motion shadow generation and its related properties are analyzed, and different kinds of motion shadow detection algorithms are introduced.The classical moving target detection algorithm is analyzed and the codebook based method is described in detail.This paper introduces the features and classifiers commonly used in target classification, and presents a motion shadow detection algorithm based on Haar Local Binary pattern feature of Hal type.The algorithm extracts the HLBP feature vectors of the moving region and the background region, and does not need to select the threshold, divide the image and statistics the histogram, and measure the texture difference using the Manhattan distance.The difference image is segmented by threshold to detect the moving shadow.Combined with color space information and texture information, a moving shadow detection method based on random forest is proposed. This method does not need to hypothesize the illumination and reflection properties of the application scene, and also avoids the setting of parameters.A random forest classifier is used to classify each pixel to determine whether it is a shadow or a foreground pixel.This method can achieve good results for various indoor and outdoor scenes and has a certain generalization ability. A moving target recognition method based on HOG-HLBP features is proposed.Using histogram of Oriented gradient histogram (histogram) to describe the contour of the target and the ability of HLBP to describe the texture, the two methods are combined to form the HOG-HLBP feature, and the support vector machine (SVM) is used to make the multi-class decision.Experiments show that this method can achieve better classification effect.A multi-target tracking method based on detection is implemented. The apparent information is added to the matching cost to enable it to deal with some occlusion cases.Combined with motion detection, motion shadow cancellation and multi-target tracking, video target classification is realized.
【学位授予单位】：中国科学技术大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】