低信噪比图像中多运动目标的实时识别研究与应用

【作者】 张杏；

【导师】 谭鹤良； 彭晓燕；

【作者基本信息】 湖南大学 ， 机械电子工程， 2002， 硕士

【摘要】 本文首先概述了模式识别的分类过程、系统设计和应用领域，分析了低信噪比图像中多运动目标实时识别的难点，重点介绍了本课题的研究内容以及解决方案。 接着，讲述了分割图像的三种结构和传统的基于边缘、边界和区域的图像分割技术以及几种常见的颜色模型。然后，分析和阐明了传统的运动目标检测方法的不足，并在此基础上结合研究中的实际实验环境，提出了一系列解决方法，包括针对降低庞大数据量而提出的网格扫描、局部“跟虫”追踪和动态窗口扫描等目标检测方法，针对实验环境中光照不均和颜色干扰提出基于人机交互的最大最小值阈值选取方法和引入改进的RGB模型到HSV模型的转换方法，为消除图像畸变对识别精度的恶劣影响而采用的通过控制点进行双线性插值进行畸变校正的方法； 紧接着，概述了神经网络的发展历史和几种常用神经网络模型的特点，重点研究了前馈型神经网络在模式识别中的应用问题，详细阐述了基本的BP算法和学习过程中BP算法的改进，从而使网络收敛速度更快，解决问题更有效，并在此基础上，设计了一个基于BP神经网络的运动目标识别系统，给出了实验结果。 机器人足球系统作为本课题研究的一个应用实例，本文在最后，介绍了机器人足球比赛的学术背景、战略意义和在现实生活以及军事上的应用前景，阐明了机器人足球比赛系统的三种控制结构，由此指出半自主型机器人足球系统是一个合适的研究平台，并分析了这种系统各个组成部分的工作原理和功能，深入讨论了整个视觉软件的实现过程，并系统地阐述了在Windows系统平台下进行图像采集卡的二次开发方法，对从事实时图像处理、图像识别和硬件二次开发的人员有一定的借鉴作用。更多还原

【Abstract】 Firstly, the process of classifying, system design and application of pattern recognition are summarized in this thesis. The difficulties in recognizing several moving objects in real-time images with low SNR are analyzed. And then the research contents as well as resolution are both explained with emphases.Next, the three saving methods for image segmentation results, the traditional technologies of image segmentation based on edge, boundary and region, together with several common color spaces are introduced. Later on, after elaborating the disadvantages of the old methods in detecting and recognizing moving objects, a series of corresponding approaches are proposed, such as grid scan, local tracking bug and dynamic window in object tracing to reduce the huge data needed to be processed, maximum and minimum for selecting a proper segmentation threshold and improved conversion from RGB model to HSV and so on to decrease the influence of inhomogeneous lighting and the color noise, a bilinear interpolation in each quadrant to eliminate the bad effect on the recognition precise because of the distortions of the camera.After that, much emphasis is given on application study in pattern recognition with a feed-forward neural network. Both the basic BP algorithm and improved BP algorithm in the study process are described in detail, and the later is used to quicken convergence speed and improve validity of the network. According to the above-mentioned introduction, the author has designed a recognition system of moving objects based on BP neural network and experimental results are showed.Finally, as RoboCup is an applied example of this project, the learning background, strategy significance and the applied foreground in our daily life and military affairs of the robotic soccer system are summarized in this thesis. After introducing of the three control constructions in robotic soccer system and comparing with each other, we pointed out the half-independent robotic soccer system is more appropriate to be ourstudy platform. The analysis of the principle and function of each component in this kind of system and the whole process of how to carry out the vision software system from the input image to the information of robots and ball and the ways in a second development of the frame grabber on the Windows system platform are both perfectly described. This will make a great help to those who do a study in real-time image processing.更多还原