基于机器视觉的移动目标跟踪系统研究

发布时间：2018-03-15 02:12

本文选题：机器视觉　切入点：目标跟踪　出处：《青岛科技大学》2017年硕士论文　论文类型：学位论文

【摘要】：机器视觉技术是近年来的研究热点,它是用机器替代人眼对目标进行检测识别和判断,而目标跟踪作为机器视觉领域的核心问题,已经广泛应用到军事类导航、智能交通管制、工业检测、医疗科学诊断以及公共生活安全等领域。尽管到目前为止人们已经提出了许多有效的目标跟踪算法,但是这些算法基本上都具有一定的针对性,即针对某一方面(高速移动、遮挡、目标变形等)提出相应的解决方案,并且大多数目标跟踪算法复杂度高,实时性低,容易发生跟踪丢失、抖动等现象,要开发出真正鲁棒、实用的跟踪应用系统还需要解决大量的问题。本文以智能婴儿车为研究平台,构建了基于机器视觉的嵌入式移动目标跟踪系统,并对所应用的跟踪算法进行深入的研究与分析。论文的主要研究内容及成果如下:1.对TLD目标跟踪算法进行了深入的分析研究,该算法将传统的跟踪算法和检测算法相结合,当目标消失后具有良好的重检测能力,并通过一种在线学习机制对跟踪结果以及检测结果的错误进行纠正,以不断提升分类器的性能,最终输出较准确的跟踪目标框,因此该算法具有良好的鲁棒性。但是通过对TLD算法中各模块的详细分析可知,跟踪模块对光照强度和遮挡等干扰比较敏感从而会影响算法的跟踪性能;检测模块需要扫描整个视频帧,并且扫描窗口需要依次通过方差分类器、集合分类器以及最近邻分类器才能检测到所要跟踪的目标的位置,这样就使算法的实时性下降。2.针对TLD目标跟踪算法实时性低的不足,提出使用基于压缩感知理论的检测模块对TLD跟踪算法进行改进,替代TLD跟踪算法中的级联分类器,以提高算法的实时性。文章根据压缩感知理论的思想基础对其基本原理进行了详细的分析研究,并通过详细分析运用测量矩阵对目标的Haar-like特征进行提取以及利用朴素贝叶斯分类器对目标和背景进行分类,证明了使用压缩感知检测机制对目标进行检测的可行性,同时证明了基于压缩感知理论的检测模块对目标的检测具有较高的实时性,从而使得将压缩感知检测模块用于TLD跟踪算法后在保证跟踪结果的精度以及鲁棒性的同时提高了算法的实时性。3.以智能婴儿车为载体设计了基于机器视觉的嵌入式移动目标跟踪系统,并将对TLD的改进算法应用到该系统中以实现婴儿车对所设定移动目标的实时跟踪,最后将改进算法的实验结果与原算法的实验结果进行比较分析,结论表明改进后的跟踪算法实时性更好,鲁棒性更高。
[Abstract]:Machine vision technology is a hot research topic in recent years. It uses machine instead of human eyes to detect and judge targets. As a core problem in machine vision field, target tracking has been widely used in military navigation and intelligent traffic control. Industrial detection, medical science diagnosis, public life safety and so on. Although many effective target tracking algorithms have been proposed so far, but these algorithms are basically targeted. That is to put forward the corresponding solution for a certain aspect (high-speed movement, occlusion, target deformation, etc.), and most target tracking algorithms have high complexity, low real-time, easy to occur tracking loss, jitter and other phenomena, to develop a real robust, The practical tracking application system also needs to solve a lot of problems. In this paper, an embedded moving target tracking system based on machine vision is built on the platform of intelligent stroller. The main contents and results of this paper are as follows: 1. The TLD target tracking algorithm is deeply studied, which combines the traditional tracking algorithm with the detection algorithm. When the target disappears, it has good re-detection ability, and corrects the tracking results and the errors of the detected results through an online learning mechanism, so as to continuously improve the performance of the classifier, and finally output a more accurate tracking target box. Therefore, the algorithm has good robustness. But through the detailed analysis of each module in the TLD algorithm, the tracking module is sensitive to the interference such as illumination intensity and occlusion, which will affect the tracking performance of the algorithm. The detection module needs to scan the whole video frame, and the scanning window needs to pass the variance classifier, the set classifier and the nearest neighbor classifier in order to detect the location of the target to be tracked. In view of the low real-time performance of TLD target tracking algorithm, an improved TLD tracking algorithm based on compression sensing theory is proposed to replace the cascade classifier in TLD tracking algorithm. In order to improve the real-time performance of the algorithm, this paper makes a detailed analysis and research on its basic principle according to the ideological basis of compression perception theory. Through the detailed analysis of using measurement matrix to extract the Haar-like feature of the target and using naive Bayesian classifier to classify the target and background, the feasibility of using compressed sensing detection mechanism to detect the target is proved. At the same time, it is proved that the detection module based on compressed sensing theory has high real-time performance for target detection. So that the compression sensing detection module can be used in TLD tracking algorithm to ensure the accuracy and robustness of the tracking results, while improving the real-time performance of the algorithm. 3. A machine vision based embedding based on intelligent baby carriages is designed. Type moving target tracking system, The improved algorithm of TLD is applied to the system to realize the real-time tracking of the moving target set by the baby carriage. Finally, the experimental results of the improved algorithm and the experimental results of the original algorithm are compared and analyzed. The results show that the improved tracking algorithm is better in real time and robust.
【学位授予单位】：青岛科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】