《智能视频监控中运动目标检测与跟踪技术研究及实现》由会员分享，可在线阅读，更多相关《智能视频监控中运动目标检测与跟踪技术研究及实现（64页珍藏版）》请在金锄头文库上搜索。
1、中南民族大学 硕士学位论文 智能视频监控中运动目标检测与跟踪技术研究及实现 姓名：贾茜 申请学位级别：硕士 专业：通信与信息系统 指导教师：谢勤岚 20090422 中南民族大学硕士学位论文 I 摘 要 基于计算机视觉的智能视频分析融合了图像处理、模式识别、人工智能、自 动控制及数学和计算机科学等多个学科领域的技术，具有重要的科学意义和广阔 的应用前景。随着信息技术的高速发展以及安全形势的迫切需要，人们对安防设 备的性能要求日益提高。利用计算机视觉技术，提高视频监控系统的自动化、智 能化是未来的发展方向。运动目标检测与跟踪是视频监控中的一个重要任务，它 是后续各种高级处理的基础，如模式识别、行
2、为分析等。现实的监控环境往往是 错综复杂、变化无常的，探求能够从容应对复杂环境的各种变化，快速、准确而 稳定的检测和跟踪目标的方法是该课题的主要研究内容。 本论文以实现智能化视频监控系统这一背景为基础，重点针对智能视频监控 系统中运动目标的检测和跟踪这两个关键技术进行研究，以期获得满足智能视频 监控更为实用的算法，以及系统整体的闭环控制方法。 论文针对运动目标检测与跟踪算法及其发展进行了全面的综述。首先，回顾 了若干常用的运动目标检测算法，包括光流法、时域差分法、背景差分法，并阐 述了基于高斯背景建模的前景检测方法；然后，对常用的跟踪算法作出了分类， 并深入讨论了 Mean Shift 理论及
3、相关的目标跟踪算法。针对这些算法，探讨了各 方法的原理，并分析、比较它们各自的优缺点和适用范围。在此基础上，论文完 成了三个方面的工作： 结合帧差法和 CAMShift 算法， 设计了一种自动的运动目标检测与跟踪算法。 首先用时间连续的三帧双差分对运动目标识别和提取，自动的选取跟踪框；再通 过 CAMShift 算法计算目标的精确位置并调整跟踪窗口大小。 将上述改进的目标检测与跟踪算法完成了在 DSP 上的实现，并设计了一个 由 DM642-PCI 开发板和伺服机云台组成的闭环控制系统。该系统首先自动检测 运动目标，再在每一帧中通过 CAMShift 算法计算目标的精确位置；最后将目标 质心与
4、视野中心的偏差信息转化成控制协议，通过串口发送，驱动云台转动来改 变摄像机视野，使目标始终可见。利用 CCS2.2 开发环境，在以 TMS320DM642 为核心的硬件平台上实现了系统的软件算法。 设计了一种高速球形摄像机 PTZ 跟踪的控制策略。球形摄像机机能在水平、 垂直、光轴三个方向运动，360范围全方位高速旋转。在球机机械参数未知的 情况下，通过控制球机做间歇性转动，调整球机 P/T 方向姿态角度使被跟踪对象 始终可见；当视野中心对准目标后，控制球机开始进行变倍动作对图像缩放，可 以观察或抓拍目标局部细节。同时，针对球机变倍控制中跟踪窗口大小自适应调 整的问题，利用 SIFT 特征匹配
5、算法设计了一种计算球机变倍率的方法，并给出 智能视频监控中运动目标检测与跟踪技术研究及实现 II 了一种在 RGB 颜色模型中基于 Mean Shift 的跟踪算法。基于 PELCO 协议，利 用 VS2005 和 OpenCV 软件平台，实现了 PTZ 跟踪的整体流程。 论文的主要研究集中在算法和系统控制流程设计上，而对它们结合硬件的实 现为实验性的工作，搭建了系统平台。为了发挥 DSP 的强大性能，还需对嵌入 式软件进一步优化，而这是个非常需要时间和工程项目经验的工作。 对于球机 PTZ 跟踪策略，通过实验证明能有效、实时的跟踪目标，并能够控制球机放大 拍摄目标的局部细节， 在犯罪取证的应
6、用场合有一定参考价值。 由于目前对于 PTZ 跟踪实现方面的参考资料较少，本文策略为预演性的研究，离真正的商业化用途 仍有一定差距。 关键词：智能视频监控；运动目标检测与跟踪；TMS320DM642；Mean Shift 算 法；高速球形摄像机；PTZ 控制 中南民族大学硕士学位论文 III Abstract Computer vision-based intelligent video analysis combines many technologies including image processing, pattern recognition, artificial intellig
7、ence, automatic control, mathematics and computer science, etc, and has vital scientific significance and broad applied prospects. Along with the fast development of information technology as well as the urgent need of security situation, the requirement for higher- powered security equipment is eve
8、r-increasing. It is becoming the developmental direction in the future that making video surveillance system be roboticized and intellectualize through computer vision technology. Both motion target detection and tracking are important tasks in video surveillance for they are the cornerstones of the
9、 following each kind of high-level processing, like pattern recognition, behavior analysis, etc. Since the realistic monitoring environment is often complicated and constantly changing, it is the main research content that searching the method which can manage various kinds of change in complex envi
10、ronment, and detect and track target rapidly, accurately and stably. Based on the background of realizing intellectualized video surveillance system, this paper mainly research on two key technologies in the intelligent video surveillance system, namely motion target detection and tracking. The purp
11、ose is to design the more practical algorithm and form closed-loop controlling method in whole to realize the intelligent video surveillance. The paper has given a comprehensive summary of the motion target detection and tracking algorithms and their development. Several commonly used motion target
12、detection algorithms are firstly reviewed, including optical flow, time-difference, background-difference, then the method of foreground detection based on the Gauss background model is elaborated. A classification of the motion target tracking algorithms is also made, and tracking algorithm based o
13、n the Mean Shift is thoroughly studied. The principles, advantages/disadvantages, and the sphere of applications of these algorithms have been analyzed and compared. With this understanding, the paper has been fulfilled following three works: An automatic tracking system has been designed by the mea
14、ns of combining differencing-frame and CAMShift. Firstly, foregroundbackground is detected through sequential three-frame-double-difference. The CAMShift Algorithm is then used to calculate the exact location of the target and adjust the size of search window. 智能视频监控中运动目标检测与跟踪技术研究及实现 IV The above-me
15、ntioned improved motion target detection and tracking method has been implemented on DSP. A closed-loop system based on DM642-PCI board and servo cradle head has been also designed. To begin with, the system detect moving target automatically, and then the CAMShift Algorithm is used to calculate the
16、 exact location of the target. The information of the warp between the targets center and the focus of the cameras eyeshot is transformed into the protocol commands and sent via the serial port lastly, to rotate the cradle head to change cameras eyeshot for maintaining the target visible in the scene. The software algorithms have implemented on the hardware with the core of TMS320DM642 by CCS2.2. A kind of tactic of PTZ tracking controlling based on high speed spherical camera ha
