基于相关相似度的在线多示例学习目标跟踪算法

Relative Similarity Based Online Multiple In-stance Learning Algorithm in Object Track-ing

目标跟踪是计算机视觉领域的研究热点之一,并得到了广泛的应用。目前许多学者将机器学习方法引入目标跟踪,例如,基于多示例学习的目标跟踪算法(即MIL)已经被提出。然而,传统的MIL跟踪算法在正负样本的选择上存在一定的不稳定性,容易在时间的推移下出现目标漂移的现象。为了克服上述问题,提出了一种简单、有效且高效的基于相关相似度的在线多示例学习目标跟踪算法。该算法通过定义相关相似度来对正包中的样本进行进一步的选择与加权,从而提高目标跟踪的性能。与新近算法的实验对比表明,本文提出的算法在目标跟踪的准确性、精度、鲁棒性等方面均有一定的提高。

Object tracking is one of the hot topics in computer vision and has wide applications. At present, many scholars have introduced machine learning methods into target tracking. For example, multiple instance learning (MIL) based object tracking has been proposed. However, the traditional MIL tracking algorithms have some instability under the selection of positive and negative samples, and they are easy to appear the phenomenon of target drifting over time. In order to overcome the above problems, this paper proposes a simple, effective and efficient object tracking algorithm using online multiple instance learning based on relative similarity. The algorithm further selects and weights the samples in the positive bag by defining the relative similarity so as to improve the performance of object tracking. By contrast to the recent algorithms, the experiments show that the algorithm in this paper has a certain increase in accuracy, precision, and the robustness of object tracking.