【作者】 郑江滨；

【导师】 赵荣椿；

【作者基本信息】 西北工业大学 ， 计算机应用， 2002， 博士

【摘要】 本论文的主要内容为智能视频监视方法的研究，其中主要涉及到在静止和运动背景下运动目标的检测和提取，运动目标的跟踪，运动目标的识别分类以及被监视图像的数字水印认证等方面的研究内容。 在静止背景下运动目标检测和提取研究方面，总结了当前三类目标检测和提取算法的原理，提出了多个改进的运动目标检测和监视算法。在介绍瞬时差分算法基本原理的基础上，提出了一个改进的瞬时差分算法，并给出了实验结果。提出了对像素进行区分的获取背景估计图像的统计平均方法；针对一个“完美”视频监视系统应该避免的问题，提出了一个能精确提取目标的算法，在对目标检测时，实现分区分级检测，实现了目标的精确提取。 在运动背景下运动目标检测和提取研究方面，证明了当背景图像慢速运动时使用一个带有六个参数的仿射运动模型来表示整个背景图像运动的合理性，提出了在慢背景运动情况下，使用背景匹配方法进行目标检测的原理。提出了用块匹配法进行背景图像匹配，该方法利用待匹配块之间仿射运动模型的变换关系，求解出了背景运动模型参数，实现了相邻两帧背景图像之间的匹配，并给出了实验结果。提出了基于的光流方程和仿射运动模型的背景运动模型参数估计方法，同时对光流的计算进行了讨论。 在运动目标跟踪研究方面，着重讨论了目标快速搜索算法。提出了使用扩展的卡尔曼滤波器预测运动目标下一时刻可能处于的区域，缩小了目标跟踪时的搜索范围，并且充分利用运动目标检测的结果，提高了目标跟踪时的匹配效率，实现了目标的快速跟踪。 在运动目标识别方面，提出了几个用于区分行人和车辆目标的特征，并在此基础上结合支撑矢量机方法实现了目标的分类，得到了较好的实验结果。 数字水印技术已经被用于数字认证，在这一方面，在讨论了数字图像水印技术特点、要求以及各种实现方法的基础上，提出了一种基于彩色分量通道DCT域的数字水印技术。该算法通过修改彩色图像两色差分量的DCT系数，将水印信号埋植到图像中，而在水印信息检测时，不需要原始图像以及其它额外的数据。经实验测试，本文提出的算法具有很高的鲁棒性和不可见性。更多还原

【Abstract】 This paper is about intelligent video surveillance. It involves in the following several topics, Moving objects detection under a static or moving background, tracking the moving objects, classifying moving objects into two class, walking man and vehicle, the last research topic is digital image watermarking for authentication.On the research of the moving objects detection and retrieval under a static background, we proposed a improved algorithm based on adjacent frame difference, and several background I image estimation methods are given, and an adaptive background estimation method by means of distinguishing moving pixels and background pixels was proposed, at end of this topic, we proposed a accurate retrieval moving objects algorithm, an experiment is given to show this method is efficient.On the research of the moving objects detection and retrieval under a moving background, we explained that using an affine-transformation model with six parameters to model the background image motion is reasonable. In order to detect moving objects, we proposed two background images matching methods. One method was based on image blocks matching, and another method is based on optic flowing estimation.On the research of the tracking objects, we researched on fast objects searching methods. In order to track moving objects real-time, we used an extend Kalman filter to predict moving objects position and then reduce objects search range, and made use of results of moving objects detection to improve objects matching efficiency, at last, an experiment was given to show that the proposed algorithm was fast and efficient.On the research of the moving objects classifying, we proposed several features that can be used for distinguish walking man and vehicle, and then we used Support Vector Machine to classify walking man and vehicle.On the research of the image authentication, digital watermarking has been proposed as a method for authentication. Now, there are many watermarking schemes for still color images, we present an approach for embedding digital watermarking information into color channels DCT (Discrete cosine transforms) domains in the image. Specifically, the embedded watermarking is extracted without resorting to original image and any other data sets. Experimental results demonstrate that the watermark is robust to most of the image and signal processing techniques, and also our schemes are shown to provide very good results both in terms of image transparency and robustness.更多还原