【作者】 胡伟；

【导师】 姚莉秀；

【作者基本信息】 上海交通大学 ， 控制科学与工程， 2014， 硕士

【摘要】 监控系统广泛用于军事、商业和交通等领域，传统的监控系统中操作人员扮演至关重要的角色，既耗费人力又不可靠，随着监控规模和数据量不断增大，对智能监控技术的需求和依赖程度也越来越大。本文利用图像处理和模式识别技术提出特定行人车辆目标的检测方法，并在此基础上研究面向多摄像头非重叠视场的人车目标匹配技术。本文研究的具体内容主要包括行人车辆颜色检测、基于单摄像头的人车目标检测和面向多摄像头非重叠视场的人车目标匹配。特定行人车辆的检测包括人车颜色检测和人车目标检测两个部分，目标检测中利用高斯混合模型进行背景建模，并结合阴影去除、形态学处理等手段来检测前景，然后提取目标的方向梯度直方图特征训练分类器；颜色是特定目标最直观的特征，颜色检测总体采用树形分类的思想，主要内容包括彩色目标和非彩色目标的区分、黑色和白色目标的分类以及彩色目标颜色的定量计算，利用直方图特征可以很好的解决前两个问题，重点在于彩色目标颜色的定量计算问题，对此，本文提出一种基于对像素点的光照估计的方法。多摄像头间的目标匹配问题是近年来的研究热点，本文针对非重叠视场中的人车目标匹配进行了研究，讨论了用于建立目标表现模型的各种特征，并提出基于多种特征的人车目标匹配方法，主要包括特征提取、相似度度量和人车匹配识别策略。更多还原

【Abstract】 Monitoring systems are widely used in military, commercial, and transportationand other areas, the traditional surveillance system operators play a crucial role, bothlabor-intensive and unreliable, as the monitor size and the amount of data is increasing,the intelligent monitoring technology demand and dependence is growing. In this paper,image processing and pattern recognition techniques proposed pedestrian trafficdetection method specific objectives and research based on multi-camera objectmatching people and vehicles based on this technology.Specific contents of this paper include pedestrian vehicle color detection, basedon a single camera people and vehicles based target detection and matchingmulti-camera target vehicles. Detect specific direction pedestrian traffic, includingvehicles and vehicle color detection and object detection in two parts, target detectionusing Gaussian mixture model for background modeling, combined with shadowremoval, morphological processing and other means to detect foreground, and thenextract the target gradient histogram feature training classifiers; colors are a particulartarget of the most intuitive features, color detection general idea of using treeclassification, the main contents include distinguishing color target and non-targetcolor, black and white target classification and color target color quantitativecalculation, histogram feature can solve the first two problems, with emphasis onquantitative target color color computing problems, this paper proposes a kind of pixelillumination estimation based approach.Multi-camera target matching problem between a research hotspot in recent years,aiming at the target vehicle matching non-overlapping field of view of the peoplestudied, discussed the establishment of performance targets for the various features ofthe model and presents a variety of features based on target vehicles matching method, including feature extraction, similarity measurement and identification of people andvehicles matching strategy.更多还原