相关性滤波器运动目标跟踪算法

发布时间：2018-04-11 11:20

本文选题：相关性滤波器 + 多尺度　；参考：《昆明理工大学》2017年硕士论文

【摘要】：运动目标跟踪是计算机视觉中的重要环节,在军用、公共安全和自动驾驶等领域有着广泛的运用。检测技术的发展早于跟踪技术,已经有许多效果突出且理论基础完善的算法。借助检测技术的实时检测跟踪(Tracking by Detection)是近年突起的一类跟踪方法,在每一帧检测到目标以实现连续视频序列目标的跟踪,具有较好的跟踪性能。其代表是Kernelized Correlation Filter算法,通过训练一个相关性滤波器进行相关性滤波以实现检测。本文的研究重点是相关性滤波器的多尺度跟踪以及模板漂移时的持续跟踪。提出一种多尺度的内核化相关性滤波器ACF算法。现有改进方法多是基于MOSSE的多尺度改进方法,本文将其扩展至KCF,利用KCF循环结构对角化的性质进行高效、高精度的位置滤波。再利用目标尺度金字塔对目标进行多尺度表达,利用卷积定理在傅里叶域中对目标进行尺度滤波,并根据上一帧检测结果的尺度变化率,适时地调用尺度优先策略或位置优先策略,使得跟踪器在权衡尺度变化与位移变化时更具鲁棒性。对于模板漂移的持久跟踪,提出一种持续目标模板——CUR滤波器。CUR滤波器以矩阵降维技术为核心,使用CUR分解中构建R矩阵的方法,从包含了所有成功检测目标信息的历史矩阵Q中构建CUR滤波器,最大程度地保留目标的特征信息,实现目标的持续性表达。当跟踪器跟踪失败时,使用CUR滤波器作为下一帧检测的目标外观模型,由于CUR具有目标持续表达的特性,因此,能够不受跟踪失败帧的影响,实现目标的持续跟踪。
[Abstract]:Moving target tracking is an important part of computer vision, which is widely used in military, public safety and autopilot fields.The development of detection technology is earlier than that of tracking technology, and there are many algorithms with outstanding effect and perfect theoretical foundation.Tracking by Detection with the help of detection technology is a kind of tracking method which has been raised in recent years. In order to realize the tracking of continuous video sequences, the target is detected in every frame, and it has good tracking performance.It is represented by Kernelized Correlation Filter algorithm, which can be detected by training a correlation filter for correlation filtering.The emphasis of this paper is the multi-scale tracking of correlation filter and the continuous tracking of template drift.A multi-scale kernel correlation filter (ACF) algorithm is proposed.Most of the existing improvement methods are based on MOSSE. In this paper, we extend them to KCFs and use the diagonalization property of KCF cyclic structure to carry out efficient and high-precision position filtering.Then the multi-scale representation of the target is performed by using the target scale pyramid, and the scale filtering is carried out in the Fourier domain by using convolution theorem, and the scale change rate of the detection result of the previous frame is calculated according to the scale change rate of the detection result.The scale-first strategy or location-first strategy is called in time, which makes the tracker more robust in balancing the scale change with the displacement change.For the persistent tracking of template drift, a method of constructing R-matrix by using CUR decomposition is proposed, which is based on matrix dimension reduction technology.The CUR filter is constructed from the history matrix Q which contains all the information of the successful target detection. The feature information of the target is preserved to the maximum extent and the persistent expression of the target is realized.When the tracker fails, the CUR filter is used as the target appearance model for the next frame detection. Because CUR has the feature of continuous expression of the target, it can realize the continuous tracking of the target without the influence of the tracking failed frame.
【学位授予单位】：昆明理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】