《基于计算机视觉的智能监控技术研究》由会员分享，可在线阅读，更多相关《基于计算机视觉的智能监控技术研究（75页珍藏版）》请在金锄头文库上搜索。
1、南京航空航天大学 硕士学位论文 基于计算机视觉的智能监控技术研究 姓名：王强光 申请学位级别：硕士 专业：机械设计及理论 指导教师：伍铁军 2011-01 南京航空航天大学硕士学位论文 I 摘 要 随着计算机视觉技术的发展，基于计算机视觉的智能监控系统得到了越来越广泛的应用， 其功能与传统的监控系统相比也有了质的变化，不仅极大的扩展了传统的监控功能，而且降低 了在监控方面投入的成本。本文在摄像机固定的情况下，对智能监控系统的关键技术进行了研 究，并实现了具有实际应用功能的智能监控系统。 运动目标的检测是智能监控系统要实现的第一步，也是至关重要的一步，检测出的运动目 标的质量将直接关系到运动目标
2、的分类和行为分析等。因此要求检测出的运动目标信息要尽可 能的完整。本文采用了常用的减背景法进行检测，并对其中关键的背景建模方法进行了研究， 对单高斯背景建模方法和混合高斯背景建模方法进行了融合，通过实验验证了该方法既可以清 晰的检测出运动目标，同时也能减少建模的时间，很好的减小外界环境的影响。 目标分类也是计算机视觉领域当中一个研究的热门问题，主要应用于智能监控系统、智能 机器人、医学图像等。通过目标分类，可以实现根据预先设定的定义来区分运动目标的种类， 从而可以进一步的进行行为分析。本文利用了目标的长宽比、占空比等四个目标特征，通过实 验统计了大量样本的均值和方差，通过最小误差概率条件下的贝
3、叶斯分类理论对监控场景内的 目标进行分类。 在运动目标的跟踪这一技术上本文使用了较为成熟的卡尔曼滤波跟踪算法，该算法可以根 据已有的运动目标的运动信息对之后的运动状态做出预测，能够很好的对其进行跟踪。同时本 文对跟踪中的难点：运动目标的遮挡问题做了深入的研究，提出了利用运动目标的面积判断遮 挡的发生时刻，通过最小二乘法对目标的轨迹进行预测，从而达到遮挡期间也尽可能不丢失目 标的效果。 通过实验证明了该方法在运动目标发生短时间遮挡时可以很好的解决跟踪丢失问题。 同时本文还对监控系统特定区域的监控等问题进行了研究。 关键词关键词：计算机视觉 背景建模 贝叶斯决策 目标跟踪 遮挡预测 基于计算机视觉
4、的智能监控技术研究 II ABSTRACT With the development of Computer Vision Technology, the intelligence video surveillance based on the computer vision has been widely applied. Compared with traditional video surveillance, its function has got the essential changed. It not only expand the original function bigge
5、st, but reduce the caption invested in the video surveillance. The key technologies in intelligence video surveillance has been searched in this paper, and also the video surveillance with practical application function has been implemented under the fixed camera. Moving object detection is the firs
6、t and most important step in the system, the quality of detected object will affect directly the classification of object and objects behaviors understanding. Therefore the information of the object must completeness as far as possible. Background subtraction is used in this paper. The method of bac
7、kground modeling which is the key has been studied. The method of single Gaussian model with mixture Gaussian model can segment the moving targets accurately.It is also able to reduce time and outside influences which disproved by experiment. Target classification is an active field of computer visi
8、on, which applied widely on many fields such as intelligence video surveillance, intelligent robot and medical image analysis. Via target classification, target can be classified into different category which predefined, and then understand objects behaviors. Based on the minimum error Bayes decisio
9、n, occupy-space-proportion, the ratio of height to width, and principal axis of inertia direction are used to classification by obtained the mean value and variance through experiment. Kalman filter is used to track the target in this paper. This algorithm can predict moving information using the ex
10、isting movement, and track accurately. The occlusion which is difficult in tracking is researched in this paper. And a new method which used targets area to confirm the point at which occlusion occurs and prediction its moving path is presented to prevent the loss. The experiment results show that t
11、his method can solve the problem which the target occurred short occlusion. And the specific area surveillance is also searched in this paper. Key Words: Computer Vision, Background Modeling, Bayes Decision, Target Tracking, Oocclusion Prediction 南京航空航天大学硕士学位论文 V 图表清单 图 1.1 智能监控系统示意图.1 图 2.1 光流法模型.9 图 2.2 帧间差分法流程图.11 图 2.3 帧间差分法实验结果.11 图 2.4 减背景法流程图.12 图 2.5 单高斯背景建模法实验结果.14 图 2.6 混合高斯背景建模方法效果实验结果.17 图 2.7 监控场景图.18 图 2.8 改进的混合高斯背景建模方法流程图.19 图 2.9 输入的监控画面.20 图 2.10 混合高斯法与本文方法实验结果对比图.20 图 2.11 中值滤波实验结果对比图.21 图 2.12 灰度变化曲线.22 图 2.13 直方图均衡化实验结果.23 图 2.14
