基于核相关滤波的视觉跟踪算法研究

发布时间：2018-07-29 16:03

【摘要】：近年来目标跟踪算法取得了显著的成果,但由于实际跟踪过程中受到复杂背景及目标的多变性等影响,目标跟踪仍具有一定的挑战性。因此,本文在核相关滤波目标跟踪算法上讨论了其优缺点并进行了研究,主要创新如下:(1)为了解决目标跟踪过程光照变化,跟踪不准确的问题,提出了一种基于视觉特性的核相关滤波跟踪方法。首先,利用颜色属性作为目标特征,从而提高跟踪器的光照不敏感性;然后,采用局部线性嵌入方法自适应降维,以达到低维特征空间;最后,根据正则化最小二乘分类器获得目标位置。实验表明,所提算法具有良好的光照不敏感性,且在复杂背景下具有更好的跟踪性能。(2)为了解决目标跟踪过程中尺度变化和遮挡的问题,提出了一种抗遮挡的目标跟踪算法。引入一个多尺度滤波器,根据滤波器的响应最大值进行尺度预测;并根据目标位置峰值尖锐度的差异性,正确更新模型。实验表明,所提算法在复杂背景下能有效地解决目标尺度变化、部分或完全遮挡等问题,综合性能有了明显的提升。(3)为了解决目标跟踪过程中快速运动、运动模糊的问题,提出了一种自适应目标响应的核相关滤波跟踪算法。在核相关滤波跟踪框架上,加入一项先验目标响应来联合优化训练分类器,使得学习的分类器可以解决循环移位造成的边界效应;当响应值小于设定阈值时,利用在线支持向量机(SVM)分类器对目标进行重定位检测,提高了算法的鲁棒性。实验表明,本文算法在目标发生快速运动、运动模糊等问题情况下,能准确、可靠地跟踪目标。本文围绕核相关滤波目标跟踪算法中面临的难点问题,展开了深入研究,并提出了相应的改进方法。实验结果表明,本文所提的三种算法分别在光照变化、遮挡、尺度变化及快速运动上,具有良好的跟踪性能。
[Abstract]:In recent years, the target tracking algorithm has achieved remarkable results. However, due to the complex background and the variability of the target in the actual tracking process, target tracking is still challenging. Therefore, in this paper, the advantages and disadvantages of the kernel correlation filter target tracking algorithm are discussed and studied. The main innovations are as follows: (1) in order to solve the problem of inaccurate tracking in the process of target tracking, A kernel correlation filter tracking method based on visual characteristics is proposed. Firstly, the color attribute is used as the target feature to improve the illumination insensitivity of the tracker. Then, the local linear embedding method is used to reduce the dimension adaptively to achieve the low dimensional feature space. The target position is obtained according to the regularized least square classifier. Experiments show that the proposed algorithm has good illumination insensitivity and better tracking performance in complex background. (2) in order to solve the problem of the change of scale and occlusion in the process of target tracking, an anti-occlusion target tracking algorithm is proposed. A multi-scale filter is introduced to predict the scale according to the maximum response of the filter, and the model is updated correctly according to the difference of the peak sharpness of the target position. Experiments show that the proposed algorithm can effectively solve the problems of target scale change, partial or complete occlusion and so on. (3) in order to solve the problem of fast motion and motion blur in the process of target tracking, the proposed algorithm can effectively solve the problems such as the change of target scale, partial or complete occlusion and so on. A kernel correlation filter tracking algorithm for adaptive target response is proposed. In the framework of kernel correlation filter tracking, a priori target response is added to optimize the training classifier, so that the learning classifier can solve the boundary effect caused by cyclic shift, and when the response value is less than the threshold value, The online support vector machine (SVM) classifier is used to detect the target location, which improves the robustness of the algorithm. The experiments show that the algorithm can track the target accurately and reliably under the condition of fast motion and motion blur. This paper focuses on the difficult problems in the kernel correlation filter target tracking algorithm, and puts forward the corresponding improvement methods. The experimental results show that the three algorithms proposed in this paper have good tracking performance on illumination variation, occlusion, scale change and fast motion.
【学位授予单位】：华侨大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】