基于光流法的目标提取与跟踪算法研究

发布时间：2018-02-05 04:09

本文关键词： 目标跟踪金字塔LK光流法光流信息 Kalman预估器 SVT　出处：《江西理工大学》2017年硕士论文　论文类型：学位论文

【摘要】：目标跟踪适用于我们生活中的方方面面,同时在工业生产中也扮演了相当紧要的角色。目标跟踪的目的在于通过视频序列不间断的定位到目标的当前位置,而难点在于跟踪过程中目标姿态、背景的不断变化、以及目标被遮挡等问题。本文主要探究将光流法应用到目标跟踪框架中的课题。光流法是目标跟踪中的一类重要算法,同时也是一种利用光流方程求解目标在图像平面上运动参数的方法。由于光流方程约束条件不足,许多基于其他假设的光流法被提了出来。本文对光流法进行了深入分析,采用了最为广泛的金字塔LK光流法为基础,发现了几点待优化之处,并提出相应的解决策略,论文主要工作如下:(1)阐述了光流法目标跟踪的定义和基本框架,指出光流法的理论基础是光流方程,并且可以分为稠密光流和稀疏光流两类方法。表明了本文采用的是基于特征点匹配的稀疏光流法,梳理了常用的特征点和光流法的原理与不足。介绍了常用的两种性能较好的光流法,同时对比了这两种光流法各自的优点与不足,说明了进一步优化相关算法的必要性。(2)针对光流匹配在遮挡时出错、LK光流法在背景和拍摄设备同时运动的复杂动态背景下目标与背景特征点难以区分等问题,结合目标跟踪框架的特点,给出了一种利用Kalman预估器和LK特征点光流信息的优化跟踪算法。首先利用光流信息定义了状态向量来筛选目标特征点;之后为了提高光流法在遮挡时的匹配精度,结合了Kalman预估器来实行优化。实验证明,优化算法在缓慢或者快速移动运动场景中的性能均表现良好,并且确保了有遮挡时跟踪精度不会降低。(3)针对基于SVM内核的SVT跟踪算法的得分不能够区分临近的同类干扰物、光流法和SVT跟踪算法对遮挡情况非常敏感的问题,给出了结合光流方程与SVT的改进方案。首先利用光流方程和SVT建立超定方程,求解得到的运动参数对同类干扰物具有鲁棒性,同时利用Kalman预估算法对运动参数进行估计,确保遮挡情况下训练器得分不会偏低。实验结果表明,在对道路上的车辆跟踪时,改进方法能够避免目标附近其他车辆的干扰,同时能够有效地提高目标车辆在部分遮挡时的跟踪精度。
[Abstract]:Target tracking is suitable for every aspect of our life and plays a very important role in industrial production. The goal of target tracking is to locate the current position of the target continuously through video sequences. The difficulty lies in the change of target attitude and background in the tracking process. This paper mainly discusses the application of optical flow method to target tracking framework. Optical flow method is an important algorithm in target tracking. At the same time, it is also a method to solve the motion parameters of the target in the image plane by using the optical flow equation, because the constraint condition of the optical flow equation is insufficient. Many optical flow methods based on other hypotheses have been proposed. In this paper, the optical flow method is deeply analyzed, and the most widely used pyramid LK optical flow method is used as the basis, and several points for optimization are found. The main work of this paper is as follows: 1) the definition and basic framework of the optical flow method for target tracking are expounded, and the theoretical basis of the optical flow method is the optical flow equation. And it can be divided into dense optical flow and sparse optical flow, which shows that this method is based on feature point matching. The principles and shortcomings of the characteristic points and optical flow methods are summarized. Two optical flow methods with good performance are introduced, and the advantages and disadvantages of the two optical flow methods are compared. It explains the necessity of further optimizing the correlation algorithm. 2) aiming at the error of optical flow matching in occlusion. The LK optical flow method is difficult to distinguish between the target and the background feature points in the complex dynamic background in which the background and the shooting equipment are moving simultaneously, which combines the characteristics of the target tracking framework. An optimal tracking algorithm based on Kalman predictor and LK feature point optical flow information is presented. Firstly, the state vector is defined to filter the target feature points. After that, in order to improve the matching accuracy of optical flow method in occlusion, the Kalman predictor is used to optimize. The experimental results show that the optimization algorithm performs well in slow or fast moving scenes. And it ensures that the tracking accuracy will not be reduced when there is occlusion.) the score of SVT tracking algorithm based on SVM kernel can not distinguish the adjacent interference. Optical flow method and SVT tracking algorithm are very sensitive to occlusion. An improved scheme combining optical flow equation with SVT is presented. Firstly, the overdetermined equation is established by using optical flow equation and SVT. The obtained motion parameters are robust to the same kind of interference, and the motion parameters are estimated by using Kalman estimation algorithm to ensure that the score of the trainer is not low under occlusion. The experimental results show that. When tracking vehicles on the road, the improved method can avoid the interference of other vehicles near the target and improve the tracking accuracy of the vehicle in partial occlusion.
【学位授予单位】：江西理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】