基于核函数的目标跟踪算法研究

发布时间：2018-05-26 19:01

本文选题：目标跟踪 + 核函数　；参考：《长安大学》2016年博士论文

【摘要】：目标跟踪是一个经典的计算机视觉问题,在机器人、监控系统和辅助驾驶等许多领域有着重要的应用。目标跟踪的任务就是跟踪视频序列中的一个目标或多个目标,其中目标可以是视频图像中任何感兴趣的对象。虽然目标跟踪技术在过去的几年里已经取得了重大的进展,但是由于目标外形变化、遮挡和背景干扰等复杂因素的影响,目前目标跟踪仍旧是一个复杂的开放性研究课题。本文研究改进了基于核函数的均值漂移(mean shift,简称MS)目标跟踪算法;并提出了一种基于核函数的扩展多通道关联滤波器(expanded multi-channel correlation filter,简称EMCCF)目标跟踪算法。论文的主要研究工作和成果如下:(1)针对目标非线性运动时的遮挡问题,提出了基于交叉-区间(cross-bin)颜色直方图和全局搜索机制的MS算法。首先利用cross-bin颜色直方图代替传统的区间-区间(bin-bin)颜色直方图表示目标特征,提高跟踪精度;当目标受到严重遮挡丢失后,通过一种尺度变化的搜索机制,在全局范围内搜索目标位置,提高抗遮挡能力。(2)针对一般的线性运动目标短时间遮挡问题,提出将两层卡尔曼滤波(kalman filter,简称KF)框架融入融合矫正背景权重直方图(corrected background-weighted histogram,简称CBWH)的MS。首先通过运动学方程建立第一层的数学模型,利用巴氏系数、滤波器噪声与跟踪结果之间的关系,自适应地调整跟踪结果,减少遮挡的影响;然后对目标模板直方图中的每个非零元素进行第二层滤波,通过动态变化的滤波残差和巴氏系数,实时调整更新滤波器中的各项参数,得到滤波后的目标模板,减少特征变化的影响。(3)提出了扩展多通道关联滤波器(expanded multi-channel correlation filter,简称EMCCF)目标跟踪算法。利用岭回归分类器和关联滤波器(correlation filter,简称CF)间的关系,在多通道关联滤波器(multi-channel correlation filter,简称MCCF)的基础上,通过核函数岭回归在时域建立EMCCF的数学模型并进行求解,解决了MCCF仅存在一种计算量大的线性最优解,只适用于离线目标检测,而不适用于在线目标跟踪的不足。
[Abstract]:Target tracking is a classic computer vision problem, and it has an important application in many fields, such as robot, monitoring system and auxiliary driving. The task of target tracking is to track one target or multiple targets in video sequence, in which the target can be the object of interest in the video image. Although the target tracking technique is over Great progress has been made in the past few years, but the target tracking is still a complex and open research topic due to the complex factors such as the change of the target shape, the occlusion and the background interference. This paper improves the target tracking algorithm based on the mean shift (MS) based on the kernel function, and proposes a basis for the target tracking. The target tracking algorithm of expanded multi-channel correlation filter (EMCCF) is the main research work and achievements in this paper as follows: (1) a MS algorithm based on cross interval (cross-bin) color histogram and global search mechanism is proposed for the occlusion problem in nonlinear motion of the target. First, the cross-bin color histogram is used to replace the traditional interval interval (Bin-bin) color histogram to express the target feature and improve the tracking accuracy. When the target is lost, the target location is searched in the global range and the anti occlusion ability is improved by a scale change search mechanism. (2) the general linear moving target is used. The problem of short time occlusion is to integrate the two layer Calman filter (Kalman filter, KF) framework into the fusion correction background weight histogram (corrected background-weighted histogram, for short). First, the mathematical model of the first layer is established through the kinematic equation, and the correlation between the filter noise and the tracking result is used. The system adaptively adjusts the tracking results to reduce the impact of the occlusion. Then, each non zero element in the target template histogram is filtered by second layers, and the parameters in the updated filter are adjusted in real time through the dynamically changing filter residuals and barson coefficients, and the filtered target template is obtained to reduce the influence of the change of the feature. (3) proposed Expanded multi-channel correlation filter (EMCCF) is extended to target tracking algorithm. Based on the relationship between the ridge regression classifier and the associated filter (correlation filter, CF), the kernel function ridge regression is used on the basis of the multi-channel association filter (multi-channel correlation filter, short for short). The mathematical model of EMCCF is established and solved in the time domain. The solution is that there is only one linear optimal solution with large computational complexity in MCCF, which is only suitable for off-line target detection, but is not suitable for the shortage of online target tracking.
【学位授予单位】：长安大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】