【作者】 杨志辉；

【导师】 黄凤岗；

【作者基本信息】 哈尔滨工程大学 ， 计算机应用技术， 2009， 硕士

【摘要】 随着人类社会的发展,社会的不安全因素也随之增多。国际上的每一次恐怖主义袭击事件都为各国的安防部门敲响了警钟。因此,许多国家越来越重视采用视频监控技术对重要部门、敏感地点、公共场所等进行监控。虽然人脸目标检测、车牌目标检测等典型目标检测方法已经日趋成熟,但是复杂环境下高可靠性的移动人体目标检测却仍然面临着很大困难,使之成为了本领域内最为重要和迫切的研究问题。同时,开展此方面的研究对目标的模式表达,以及本领域内的核心问题的研究也具有重要的理论意义。目前对于特征的描述主要分为对色彩的处理和对轮廓的提取两方面。其中最具有代表性的就是Viola提出的Haar-Like和Dalal采用的HOG作为物体的特征描述子,而且在人体检测上都达到了很好的效果。本文作者受到以上两种特征描述子的启发,提出了一种新的特征—多尺度方向特征,这种新特征不仅囊括了上面两种特征各自的优点,还弥补了它们的不足。这种特征是针对特征区域的形状进行统计的,由“全部特征组合起来所形成的特征集”反映的是“图像在不同尺度上的方向特征”。并且分别在SVM和AdaBoost机制下进行训练,利用训练出来的模型在视频和图片上进行行人检测,通过在公共测试集和本文自己的测试集上进行测试,并将结果与国际上其它领先算法相对比分析,实验证明了:在相同的检测标准下,使用本文所提出的理论框架,无论在运算速度上,还是在检测结果的精度上,本文算法都表现出了明显的优势。更多还原

【Abstract】 With the development of human society , there are more and more social insecurity factors at the same time. Each terrorist attack on the international gives a wake-up call for every national security departments. As a result, many countries pay more and more attentions to the use of video surveillance technology to the important sectors, sensitive locations, such as to monitor public places. Although the object detection technology such as face detection and license plate detection have become more mature, however the high reliability detection of moving human which is on the complex environment is still faced with great difficulties. Meanwhile, to carry out the research also has important theoretical significance to the objective model expression as well as the core of this area.At the present, the description of the main features contains tow parts which are the processing of color and the extraction of the outline. The most representative descriptors are Viola’s Haar-Like and Dalai’s HOG., and achieve good results in human detection. In this thesis, we put forward a new descriptor-Multi-scale orientation feature. The new features not only include the two characteristics of their respective advantages, but also make up for their lack. This feature is described according to the shape of the characteristics region, and the set of all the features describe the different orientation of each scale in a picture. Training these features in SVM and AdaBoost separately and use the training result to detect the human in video and picture. According to the test on the public testing sets and our own testing sets in this thesis, with the analysis of the result comparing to other international algorithms, it can proves that: on the same condition of detection principles, our algorithms have shown a clear advantage both on the computing speed and the accuracy of detection result.更多还原