基于核相关滤波器的目标跟踪方法研究

发布时间：2018-08-21 08:37

【摘要】：近年来,随着科技的发展和社会的进步,计算机视觉技术的应用已越来越多地呈现在人们的生活当中,其中有大部分的应用都是基于视频监控的应用,学校、医院、车站等公共场所都有大量的监控摄像头,为人们的社会生活安全提供了保障,此外交通监控也变得越来越智能化,公路、铁路等部署的智能探头已经能够自动解决很多事宜,并能记录关键的事件,上述智能监控应用的技术实现多数依靠计算机视觉理论,而计算机视觉分支之一目标跟踪也成为近年的研究热点,但是真实场景中,目标跟踪面临很多困难,具有遮挡、目标多尺度变化、光照变化等阻碍因素,而本文工作的目的就是在前人的工作基础之上进行研究,主要解决目标跟踪过程中的遮挡和尺度变化两个问题,具体工作如下:(1)本文提出一种基于核相关滤波器KCF(Kernelized Correlation Filters)的目标尺度估计以及处理遮挡的跟踪方法SOD-KCF(Scale Estimation and Occlusion Detection-KCF)。目标尺度估计模型对当前目标进行多尺度化,通过计算下一时刻各个尺度对应的分类器响应最大值构成一个序列,选取最大值并将其对应的尺度作为当前目标跟踪尺度。本文根据前向分类器响应最大值的分布特征建立遮挡处理模型,其采用阈值方法进行遮挡检测,在目标受到遮挡之后通过块区域螺旋搜索方法进行目标搜索,在目标搜索过程中计算滑动框的响应判定是否为目标。本文算法在OTB(Object Tracking Benchmark)测试序列集上测试并与4种跟踪算法进行对比实验,在跟踪准确度与跟踪成功率上分别比次优方法提高了6.1%和1.5%。(2)为了进一步解决尺度变化和部分遮挡问题,本文在核相关滤波跟踪算法基础之上提出一种尺度空间滤波器以及分块处理的方法MSKCF(Multi-Block and Scale Space-KCF)。此外,为了提高算法的鲁棒性,本文建立一种外观更新模型,该模型能够近似目标遮挡和形变的变化状态。此外,本文提出了一种自适应更新学习率的模型,取代核相关滤波器算法原本固定的学习率,加强了跟踪算法处理每个子块的鲁棒性。实验结果表明,本文算法在OTB测试序列集上优于其他对比的算法,特别地,在遮挡与尺度变化的测试序列中与核相关滤波跟踪算法对比,分别有近8%和18%的性能提高。因此,本文算法对原有算法的改进具有准确性和有效性。
[Abstract]:In recent years, with the development of science and technology and the progress of society, the application of computer vision technology has been more and more popular in people's lives, most of which are based on video surveillance applications, schools, hospitals, Stations and other public places have a large number of surveillance cameras, providing security for people's social life. In addition, traffic monitoring has become more and more intelligent. Intelligent probes such as roads and railways have been able to solve many problems automatically. And can record the key events, most of the technology of intelligent monitoring application depends on computer vision theory, and one of the branches of computer vision target tracking has become a hot spot in recent years, but in the real scene, There are many difficulties in target tracking, such as occlusion, multi-scale change of target, illumination change and so on. The purpose of this paper is to study on the basis of previous work. This paper mainly solves the two problems of occlusion and scale change in the process of target tracking. The main work is as follows: (1) this paper proposes a target scale estimation based on Kernel correlation filter (KCF (Kernelized Correlation Filters) and a tracking method (SOD-KCF (Scale Estimation and Occlusion Detection-KCF) to deal with occlusion. The target scale estimation model carries on the multi-scale to the current target. By calculating the maximum value of the classifier response corresponding to each scale at the next time, a sequence is formed, and the maximum value is selected and the corresponding scale is taken as the current target tracking scale. In this paper, the occlusion processing model is established according to the distribution characteristics of the maximum response of the forward classifier. The occlusion detection is carried out by the threshold method, and the target search is carried out by the block region spiral search method after the target is occluded. In the process of target search, the response of the sliding frame is calculated to determine whether the target is the target. This algorithm is tested on the OTB (Object Tracking Benchmark) test sequence set and compared with four tracking algorithms. The tracking accuracy and the tracking success rate are improved by 6.1% and 1.5% respectively. (2) in order to solve the problem of scale variation and partial occlusion further, Based on the kernel correlation filter tracking algorithm, a scale-space filter and block processing method MSKCF (Multi-Block and Scale Space-KCF) is proposed in this paper. In addition, in order to improve the robustness of the algorithm, a new appearance updating model is established, which can approximate the changing state of the object occlusion and deformation. In addition, this paper proposes an adaptive update learning rate model, which replaces the original fixed learning rate of the kernel correlation filter algorithm, and enhances the robustness of the tracking algorithm to deal with each sub-block. The experimental results show that the proposed algorithm is superior to other contrast algorithms in OTB test sequence sets. In particular, the performance of the proposed algorithm is improved by nearly 8% and 18%, respectively, compared with the kernel correlation filter tracking algorithm in the test sequence of occlusion and scale variation. Therefore, this algorithm is accurate and effective to improve the original algorithm.
【学位授予单位】：浙江理工大学
【学位级别】：硕士
【学位授予年份】：2018
【分类号】：TP391.41;TN713

【相似文献】