基于特征的运动目标视频检测与跟踪方法研究

发布时间：2018-08-25 18:21

【摘要】：在计算机视觉的不断发展中,视频运动目标检测与跟踪逐渐被人们所重视,逐渐成为其中重要研究方向之一。它涉及到线性代数,随机数学,图像处理,数理分析等多门学科的相关内容,无论是在军工,民用,商用方面,还是在安全防护,交通管理,智能识别等领域内,都得到了广泛的应用。已经初步实现了用计算机代替人眼,智能分析外界物体的运动,替人们对收集到的视频信息作出判断。本文介绍了研究视频运动目标检测与跟踪的课题的目的和意义以及该研究国内外主要的现状。在运动目标检测中介绍与分析了背景差分法、帧间差分法、光流法三种常用算法。为了更好地进行目标检测与识别,需先对视频图像进行去噪处理。对于当前常用的多种背景建模方法,本文进行横向与纵向的对比分析,找出尽可能优秀的算法,最终决定采用混合高斯背景建模方法完成目标检测。在运动目标跟踪方面,本文采用Meanshift算法进行跟踪。先建立目标模型,对目标的特征进行提取与分析,选取具有代表性的特征集,在接下来的每帧图像中,在以上一帧目标位置为中心,一定范围内的与该特征值相似的点,计算其Meanshift向量,向量指向的位置就是与特征最为相似的位置,从而通过特征位置判断目标所在区域。所以在Meanshift算法中,特征的提取与相似度判断尤为重要。针对传统meanshift算法中,对于高速或高变速运动的目标常发生跟踪失败的情况,以及固定的跟踪窗口大小不适合目标在视频中大小发生巨大改变的情况。本文采用了在算法中加入自适应调整,引入目标整体运动方向与速度的记录与预测,优化了传统meanshift算法对高速与高变速运动目标的跟踪的精确度。同时,对meanshift算法中核函数进行处理,通过更改核函数宽度的大小,来实时改变跟踪窗口的大小,使之更为适应目标在当前视频图像中的大小,提高了算法的精确度。
[Abstract]:With the development of computer vision, video moving target detection and tracking have been paid more and more attention and become one of the important research directions. It involves linear algebra, stochastic mathematics, image processing, mathematical analysis, and so on, whether in the military, civil, commercial, safety protection, traffic management, intelligent identification, and so on. Are widely used. It has been preliminarily realized to use computer instead of human eyes to intelligently analyze the motion of external objects and to judge the video information collected. The purpose and significance of the research on video moving target detection and tracking are introduced in this paper. This paper introduces and analyzes three common algorithms for moving target detection: background differential method, inter-frame differential method and optical flow method. In order to detect and recognize the target better, the video image should be de-noised first. For the current common background modeling methods, this paper makes a comparative analysis between horizontal and vertical, finds out the best algorithm, and finally decides to use mixed Gao Si background modeling method to complete target detection. In the aspect of moving target tracking, Meanshift algorithm is used to track. Firstly, the target model is established, the feature of the target is extracted and analyzed, and the representative feature set is selected. In each frame image, the target position of the previous frame is the center, and the point similar to the characteristic value is found in a certain range. The Meanshift vector is calculated, and the vector points to the position that is the most similar to the feature, so that the location of the target is judged by the feature position. So in Meanshift algorithm, feature extraction and similarity judgment are very important. In the traditional meanshift algorithm, the tracking failure often occurs for the high speed or high speed moving target, and the fixed tracking window size is not suitable for the large change of the target size in the video. In this paper, the adaptive adjustment is added to the algorithm, and the recording and prediction of the whole moving direction and velocity of the target are introduced. The accuracy of the traditional meanshift algorithm for tracking the high speed and high variable speed moving targets is optimized. At the same time, the kernel function in the meanshift algorithm is processed. By changing the width of the kernel function, the size of the tracking window can be changed in real time, which makes it more suitable for the size of the target in the current video image and improves the accuracy of the algorithm.
【学位授予单位】：中国科学院大学(中国科学院长春光学精密机械与物理研究所)
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】