抗模糊抗局部低照度图像特征描述子和监控相机接力邻接拓扑结构的建立是视频监控中的应用基础和研究热点。视频监控中，最基本的应用是目标检测和目标匹配，它们都是基于特征匹配的。局部特征可以很大程度地克服全局特征易受到背景杂乱及目标遮挡等因素影响的缺点，因而获得广泛使用。由于不同时间、不同地点，以不同视角和不同监控相机，再加上光照及遮挡等因素的影响，获取的同一目标的图像间往往差别很大，有的局部区域亮度特别低，有的比较模糊，直接利用SIFT(Scale Invariant Feature Transform)和SURF(Speeded Up Robust Features)等局部特征描述子，无法从中提取到较多的和质量较高的特征点，给图像间的鲁棒匹配带来了困难。监控相机海量部署使得监控相机接力关系的确立更需要自动化和智能化，人工管理已无法适应监控相机日益增长的规模。因此，进行抗模糊抗局部低照度图像特征描述子和监控相机接力邻接关系研究，具有重要的理论意义和应用价值。 就抗模糊抗局部低照度图像特征描述子和监控相机接力邻接关系进行研究，提出了一种模型和三种新算法，并通过理论和实验进行了分析与验证。 首先，提出了一种符合人类视觉系统的自底向上再到自顶向下的视觉认知过程的图像局部特征描述子的提取模型BUTD(Bottom-Up to Top-Down)。该模型解决了视频监控中遇到的从低对比度、模糊和细节不清晰等图像中提取特征点数量少和质量低的问题。该模型通过研究并借鉴了现有的由客观内容驱动的自底向上视觉注意模型和由主观命令指导的自顶向下视觉注意模型的思想而提出。该模型通过自底向上的视觉注意模型得到原始图像的总显著图，再采用自顶向下的视觉注意模型从总显著图中提取图像局部特征描述子，以用于图像的匹配和目标识别。该模型解决了目前自顶向下的视觉注意模型复杂性较高的问题，比现有模型更适用于从模糊图像、光滑图像和局部低对比度图像中提取局部特征点。 然后，提出了一种新的抗模糊的图像局部特征描述子FSIF(Fuzzy and Scale Invariant Feature Descriptor)。该描述子解决了SIFT从模糊图像中提取特征点少的问题。该描述子利用局部二值模式描述子LBP(Local Binary Pattern)的光照不变性和高独特性的优点，将LBP描述子引入SIFT描述子的构建过程中，以局部二值模式高斯尺度金字塔代替SIFT的高斯尺度金字塔，来检查FSIF的特征点。对FSIF特征点的精确提取所依赖的对比度阈值和主曲率阈值参数做了优化。该描述子对光照条件的变化具有更强的健壮性，对模糊的目标图像能准确提取到较多的特征点，显著提高了对模糊图像的匹配率。 其次，提出了一种抗局部低照度的图像局部特征描述子ALLI(Anti-Local Low Illumination Feature Descriptor)。该描述子解决了从局部低对比度图像中提取特征点的数量少和质量低的问题。ALLI描述子利用图像的相位信息具有局部对比度不变性并能包含图像中的角、线、纹理和轮廓、结构等优势，先提取原始图像的相位一致性图像，再在相位一致性图像中使用类似SURF算法的方法提取特征点和特征描述子。先计算两个特征点间的欧式距离来进行初次匹配，再利用M估计抽样一致性对误匹配的特征点对进行剔除以实现精确匹配。ALLI算法比现有描述子更适用于局部低对比度图像的匹配，提高了局部低对比度图像的匹配效果。 最后，提出了GIS(Geographic Information System，地理信息系统)和GPS(Global Positioning System，全球定位系统)的监控相机邻接关系算法。该算法解决了视频监控网络中海量监控相机接力调度自动化问题。该算法利用GIS和新加入监控相机的GPS信息，结合地理信息系统的空间索引，根据各监控相机经纬度信息，利用排序算法自动生成和其接力邻接相机的监控接力邻接关系。本算法简单且易于实现，可用于城市、公路交通道路等监控环境中的异常行为自动实时跟踪或事后自动查找逃逸路线等监控相机接力拓扑结构的构建，为改善目前完全依靠人工事后查看监控图像的工作量大、效率低的工作方式以及监控相机的海量部署、监控系统的实时接力响应提供技术基础，具有极大的应用价值和社会效益。 关键词：图像匹配，特征描述子，特征点匹配，目标识别，接力跟踪，路网拓扑

 Anti-blur and anti-local low illumination image feature descriptors and the construction of relay adjacency topology of surveillance cameras are the application basis and research focuses in video surveillance. In video surveillance, the basic applications are target detection and target matching, which are based on feature matching. Local features can be widely used because they can largely abandon the shortcomings of the global features that are susceptible to background clutter and occlusion. Due to different time and different location, different viewing angles and different surveillance cameras, illumination changes and occlusions, the obtained images of the same target are often very different. Some local areas are less bright and some local areas are blurred. When directly using local feature descriptors such as SIFT (Scale Invariant Feature Transform) and SURF (Speeded Up Robust Features), it is impossible to extract more or higher quality feature points, which brings about difficulties to robust matching between images. The massive deployment of surveillance cameras requires the establishment of camera relay relationships more intelligent since manual management is unable to adapt to the large size of surveillance cameras. Therefore, it is of great theoretical and practical significance to research on anti-blur and anti-local low illumination image feature descriptors as well as on the construction of relay adjacency topology of surveillance camera. The research is carried out on anti-blur and anti-local low illumination image feature descriptors and the construction of relay adjacency topology of surveillance cameras. One new model and three new algorithms are proposed, and they are analyzed and verified with theoretical illustration and experimental demonstration. Firstly, in order to solve the problems of little feature points and feature points with low quality that are extracted from images with local low contrast, blurred part and unclear details in video surveillance, the Bottom-Up and Top-Down visual attention models are studied. Combining with the Bottom-Up visual attention model driven by objective content and the Top-Down visual attention model guided by subjective command, the local feature descriptor extraction model of the image is proposed. It satisfies the visual cognitive process from Bottom-Up to Top-Down in human visual system. The model obtains the total saliency map of the original image through the Bottom-Up visual attention model, and then the image local feature descriptor is extracted from the total saliency map for image matching and target recognition by using the Top-Down visual attention model. The proposed model solves the problem of the difficulty and complexity of the current Top-Down visual attention model, and it is more suitable for extracting local feature points from blurred images, smooth images and local low-contrast images than other existing models. Secondly, in order to solve the problem of extracting too fewer feature points from fuzzy images, a new anti-blur image local feature descriptor FSIF (Fuzzy and Scale Invariant Feature Descriptor) is proposed. The descriptor takes the advantages of illumination invariance and high distinctiveness of LBP (Local Binary Pattern) and introduces the LBP descriptor into the construction process of the SIFT descriptor. The Gaussian scale pyramid of the SIFT descriptor is replaced by the local binary pattern Gaussian scale pyramid in order to extract the feature points of the FSIF. The parameters of contrast threshold and principal curvature threshold of FSIF descriptor are optimized in order to extract more accurate feature points. The descriptor has stronger robustness to the changes of illumination. It can accurately extract more feature points from the blurred target image and significantly improve the matching rate of the blurred images. Then, in order to solve the problem of extracting too fewer of feature points and feature points with low quality from local low-contrast images, an ALLI (Anti-Local Low Illumination Feature Descriptor) is proposed. The ALLI descriptor takes the advantages that the image phase information is local contrast invariant and it can include the angles, lines, texture and contour, structure in the image. First, the phase consistency image of the original image is extracted, and then the feature points and feature descriptors are extracted from the phase consistency image by using the similar method with SURF algorithm. The Euclidean distance between two feature points is calculated to perform the initial matching, and then the MSAC(M-estimator SAmple Consensus) is used to eliminate the mismatched feature point pairs to achieve accurate matching. The ALLI algorithm is more suitable for matching local low-contrast images than the existing descriptors, and it can improve the matching and recognition between local low-contrast images. Finally, being aimed at the automation issue of video camera relay scheduling in large scale video camera surveillance network, a video surveillance camera adjacency relation algorithm is proposed by using GIS (Geographic Information System) and GPS (Global Positioning System). When a new camera is to be deployed in a section of a road, the related cameras deployed originally in the same section are searched by using the spatial index of GIS based on its GPS information, and then the adjacency relationship topology between the new camera and the related ones is generated by using the sorting algorithm according to their GPS information. The algorithm is simple in operation and it is easy to be implemented in the real world. Specifically, it can be used in the construction of camera relay-surveillance topology, such as automatic real-time tracking on abnormal behavior and the analysis of the escape routes in city streets and other traffic roads. It can improve the current working mode that relies on humans to observe and monitor images during and afterwards, which is challenging in workload and low in efficiency. It provides a technical basis for massive deployment of cameras and real-time relay response of monitoring systems, with great economic and social benefits. Keywords: feature descriptor, feature point matching, image matching, object recognition, relay tracking, road network topology