多摄像机连续目标跟踪系统的应用研究

发布时间：2018-03-15 08:13

本文选题：多摄像机跟踪　切入点：目标交接　出处：《聊城大学》2015年硕士论文　论文类型：学位论文

【摘要】：基于视频监控的运动目标跟踪技术属于人工智能的分支,涉及到模式识别、图像处理、计算机视觉、机器学习等学科。目前,单摄像机的目标检测和跟踪技术发展的比较成熟,取得了大量的研究成果。但是,仍有一系列的难题需要解决。比如,目标之间出现遮挡、复杂背景的干扰、动态背景情况的跟踪等。多摄像机目标跟踪是近年来的研究热点,而目标交接是实现目标匹配和场景切换跟踪的关键技术,是多摄像机监控必须解决的难题。本课题在研究单摄像机运动目标检测、跟踪算法基础上提出了新的方法,并重点研究了双摄像视野重叠情况下的视野分界线生成和目标交接算法,主要工作如下:1.采用了融合的目标检测算法。首先采用帧间差分判断出目标所在位置的大致范围,并锁定中心位置,然后在此范围附近,采用背景差分的方法,准确判断目标区域。采用此种方法的优点是,对目标区域的判断准确,且受外界扰动的影响较小,可以得到较为完整的目标前景区域。2.采用了粒子滤波改进的Camshift目标跟踪算法,结合Camshift算法对粒子滤波算法进行改进。同时解决了粒子滤波需要计算大量粒子,粒子收敛速度慢,跟踪实时性差问题和遮挡情况下Camshift算法目标跟踪鲁棒性低问题。3.提出了基于SURF快速特征点匹配和投影不变量的视野分界线生成算法。首先利用SURF算法快速获得相邻两个摄像机中带有重叠区域的背景图像的特征点,然后利用RANSAC算法对特征点向量进行优化,去除匹配误差较大的点。通过对RANSAC算法的迭代,选择出4对最佳匹配点,最后,根据两幅背景图像的边界点以及投影不变量求两幅图像中重叠区域,并进行标定,生成视野分界线。4.提出了基于视野分界线几何坐标变换曲线匹配方法的目标交接算法。首先实时检测公共视野区域内是否有运动目标出现,然后进行坐标变换,并在新的坐标系下,对目标像素进行统计,最后,将得到的像素统计曲线进行匹配和标定。
[Abstract]:The branch target video surveillance tracking technology is based on artificial intelligence, involves pattern recognition, image processing, computer vision, machine learning and other disciplines. At present, the target detection and tracking technology development of single camera is relatively mature, made a lot of research results. However, there are still a series of problems need to be solved. For example between, target occlusion, complex background and dynamic background tracking. Multi camera tracking is a hot research topic in recent years, but the goal is to achieve the goal of key technology transfer, and scene change tracking, multi camera surveillance is a difficult problem to be solved. This topic in the study of single camera moving target detection, tracking the algorithm is proposed based on the new method, and focuses on the dual camera view overlap under the vision line generation and target handoff algorithm, the main work is as follows 1.: the target detection algorithm fusion. Firstly, using frame difference estimate approximate range of target position, and the locking range near the center position, and then, using the background difference method, accurately determine the target area. The advantages of this method is that the target area is accurate, and less influence by the external disturbance, the target tracking algorithm can get more complete foreground.2. using the improved particle filter Camshift, combined with Camshift algorithm to improve particle filter algorithm. At the same time to solve the particle filter need to calculate a large number of particles, the particles slow convergence of Camshift algorithm and target occlusion tracking real-time robust tracking problems the problem of low.3. proposed SURF fast feature point matching and projective invariants based on the view of demarcation line generation algorithm. SURF algorithm is first used to obtain fast The background image feature points with overlap area of two adjacent the camera, and then use RANSAC algorithm to optimize feature vector, remove the matching error larger. Through the iteration of the RANSAC algorithm, selected 4 of the best matching point, finally, according to the overlap area of the two painting like boundary points and projective invariants for the two images, and calibrate the generated vision line.4. proposed target handoff algorithm vision line geometric coordinate transformation curve matching method based on real-time detection. The first public view area if there is a moving target, and then the coordinate transformation, and in the new coordinate system, statistics, on target pixel finally, the statistics of pixel curve obtained by matching and calibration.

【学位授予单位】：聊城大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TP391.41;TN948.6

【参考文献】