基于局部差别性分析的目标跟踪算法

在复杂场景下，为了更好地提升跟踪的鲁棒性，基于局部的相似度测量得到了广泛应用。然而，局部遮挡，形变和光照变化等场景的复杂性，基于传统局部相似度测量的目标跟踪存在很大缺点，例如，在跟踪过程中，仅仅依靠目标和模板的匹配度容易造成跟踪的偏移现象。鉴于此，该文提出一种基于局部差别性相似度测量的目标跟踪算法。首先，以目标-背景的差异性，形成相似性和差异性相结合的局部判别性相似度测量；其次，基于子块在视频序列中的差异性，对子块进行差异性学习，以提高跟踪的准确性。最后，在粒子滤波框架下，基于差别性局部区域测量构建了一种有效的目标跟踪算法。实验结果表明，在复杂图像序列中，该算法实现了目标的准确跟踪，并在光照变化、旋转、缩放和遮挡等方面具有较好的效果。

Robust Object Tracking Based on Local Discriminative Analysis

The local similarity measurements are usually used for improving the tracking robustness under the complex scene. However, this method have drawbacks in cases of partial occlusion, deformation and rotation. For example, the method only considers traditional similarity measurements of targets and templates, results in the matching errors to lead to tracking failure. In this paper, a target tracking algorithm is proposed based on measurements of the local difference similarities. The presented method has the following advantages: firstly, both similarities and differences are considered for measurement; secondly, the differential weight learning of the local region is carried out to improve the accuracy of sub-block difference measurement; at last, an effective and efficient tracker is designed based on the difference analysis and a simple update manner within the particle filter framework. Experimental results show that the proposed algorithm achieves better performance than traditional competing methods in various factors, such as illumination changes, part occlusion, scale changes and so on.