Abstract:

In correlation filter-based tracking methods, redundant dense samples generated by cyclic shifts suffer from both boundary effect and unbalanced categories, which mainly upper bound the training performance. To figure out the two above problems, a margin-maximized correlation filter advanced by mask mechanism is proposed, where the purified samples can be trained in a more effective way. A dual variable is used to convert the original margin maximization problem into a loss minimization one. The final advanced correlation filter can be iteratively solved out by the alternating direction methods of multipliers. Plenty of experimental results indicate that the proposed algorithm can significantly improve the correlation filter. Compared with other tracking algorithms, the proposed tracking algorithm has obvious advantages in speed and location accuracy.

Key words: object tracking, correlation filter, boundary effect, margin maximization

摘要：

相关滤波跟踪方法中经由循环移位得到的密集样本一方面会包含有非真实的边界，另一方面样本会存在冗余性且正负比例不平衡，这两点因素很大程度上降低了相关滤波器的训练效果。为解决以上问题，基于最大化间隔思想引入了一种掩膜机制，使得滤波器能够更有效地学习净化过的正负样本。在求解目标函数的过程中，先利用对偶变量将间隔最大化问题转换成损失最小化问题，再通过交叉方向乘子法交叉迭代求出最终的相关滤波器。实验结果表明，间隔最大化与掩膜机制结合后能够有效地提升相关滤波器的跟踪性能，且相比其他跟踪算法具有明显的速度与定位精确度优势。

关键词: 目标跟踪, 相关滤波, 边界效应, 最大化间隔