基于相关滤波的目标跟踪研究

发布时间：2018-03-25 19:52

本文选题：目标跟踪　切入点：相关滤波　出处：《安徽大学》2017年硕士论文

【摘要】：作为计算机视觉领域最具挑战的关键技术之一,目标跟踪在视频监控、导航、军事、人机交互、虚拟现实、智能机器人、自动驾驶等多个领域都有着广泛的应用。经过三十多年的研究,目标跟踪领域相继涌现了大量经典、优秀的跟踪算法,但受限于现实环境以及目标运动的复杂性,当前的跟踪算法在准确性、鲁棒性以及实时性上难以满足实际的应用需求。准确、鲁棒、高效的目标跟踪算法仍然是极具挑战的研究课题。相关滤波跟踪(correlationtracking)自提出以来,其就以兼顾准确性和速度的优势,吸引了大量研究者的关注。相关滤波器通过傅里叶变换将滤波器操作转换到频域,极大提升了算法运行速度,实现了目标位置中心的快速检测,并且其重新采样在线更新滤波器,保证了算法的准确度和实时性。本文深入研究了基于相关滤波的目标跟踪算法,针对特征融合、尺度估计以及滤波器更新策略进行改进,并在此基础之上,结合相关滤波器跟踪状态判断和级联目标检测,实现了稳定的长时间目标跟踪。本文主要的研究内容和创新点总结如下:(1)首先介绍了目标跟踪领域的研究背景与意义、发展现状以及技术挑战,并归纳总结当前目标跟踪领域主流的算法框架,然后概述相关滤波器的基本概念及其在目标跟踪上的应用原理。(2)为了提高相关滤波跟踪的精度和成功率,提出了基于相关滤波的尺度和学习率自适应跟踪算法。首先算法融合了高效的特征提取方法作为滤波器输入目标样本的外观表示;针对相关滤波器不能应对目标尺度变化的限制,结合光流跟踪的思路,根据相邻帧之间可靠关键点的位移变化估计目标尺度;并采用学习率自适应方法,改进相关滤波器的更新策略。通过高效的特征提取、尺度估计以及学习率自适应方法的综合运用,大幅提升了跟踪准确度,同时也相对节省了算法运算量,保证跟踪器的实时性。在ObjectTrackingBenchmark上进行算法的对比实验、成分分析实验以及定性评估实验,以验证算法改进的有效性。(3)针对长时间跟踪过程中面临的难题和挑战,提出了基于相关滤波和级联检测的长时间目标跟踪算法。首先采用基于相关滤波跟踪的改进算法作为基础跟踪器,并结合跟踪目标状态判断、级联检测丢失目标的策略,组成长时间跟踪的算法框架。其中级联检测器分别包括基于颜色模型的局部检测、最近邻检测以及微调三个模块,高效的跟踪目标状态判断方法则是能够及时启动级联检测器的关键。通过级联检测逐层筛选搜索样本,找回丢失的跟踪目标,提升了算法在长时间跟踪中的稳定性。
[Abstract]:As one of the most challenging key technologies in the field of computer vision, target tracking in video surveillance, navigation, military, human-computer interaction, virtual reality, intelligent robot, Autopilot and other fields have been widely used. After more than 30 years of research, a large number of classic and excellent tracking algorithms have emerged in the field of target tracking, but limited by the real environment and the complexity of target motion. The current tracking algorithms are difficult to meet the practical application requirements in accuracy, robustness and real-time. Accurate, robust and efficient target tracking algorithm is still a challenging research topic. With the advantage of both accuracy and speed, it has attracted the attention of a large number of researchers. The correlation filter transforms the filter operation into frequency domain by Fourier transform, which greatly improves the speed of the algorithm. The fast detection of target location center is realized, and its resampling online update filter ensures the accuracy and real-time of the algorithm. In this paper, the target tracking algorithm based on correlation filter is studied in depth, aiming at feature fusion. Scale estimation and filter update strategy are improved, and based on this, correlation filter tracking state judgment and cascade target detection are combined. The main research contents and innovations of this paper are summarized as follows: first, the background and significance of the research in the field of target tracking, the current situation of development and the technical challenges are introduced. Then the basic concept of correlation filter and its application in target tracking are summarized. In order to improve the accuracy and success rate of correlation filter tracking, the paper summarizes the main algorithm framework in the field of target tracking, and then summarizes the basic concept of correlation filter and its application in target tracking. The scale and learning rate adaptive tracking algorithm based on correlation filter is proposed. Firstly, the efficient feature extraction method is used as the appearance representation of the filter input target sample. In view of the fact that the correlation filter can not cope with the limitation of the change of target scale, combined with the idea of optical flow tracking, the target scale is estimated according to the displacement change of reliable key points between adjacent frames, and the learning rate adaptive method is adopted. The updating strategy of correlation filter is improved. By using efficient feature extraction, scale estimation and adaptive learning rate method, the tracking accuracy is greatly improved, and the computational complexity of the algorithm is also saved. In order to verify the effectiveness of the improved algorithm, the real-time performance of the tracker is verified by the contrast experiment of algorithm, component analysis experiment and qualitative evaluation experiment on ObjectTrackingBenchmark to solve the problems and challenges in the long time tracking process. A long time target tracking algorithm based on correlation filtering and concatenated detection is proposed. Firstly, the improved algorithm based on correlation filter is used as the basic tracker. The cascade detector consists of three modules: local detection based on color model, nearest neighbor detection and fine tuning. The efficient tracking target state judgment method is the key to start the cascade detector in time. Through cascading detection to filter the search samples layer by layer, the missing target can be retrieved, and the stability of the algorithm in the long time tracking is improved.
【学位授予单位】：安徽大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】