Correlation-Filter Based Scale-Adaptive Visual Tracking With Hybrid-Scheme Sample Learning
Correlation-Filter Based Scale-Adaptive Visual Tracking With Hybrid-Scheme Sample Learning
复制标题
DOI：
10.1109/access.2017.2759583
复制
发表时间：
2018
期刊：
影响因子：
3.9
通讯作者：
Wenhui Huang;J. Gu;Xin Ma;Yibin Li
中科院分区：
文献类型：
--
作者：
Wenhui Huang;J. Gu;Xin Ma;Yibin Li
在视觉跟踪中，成熟的尺度估计方法可以大大提高跟踪性能，为模型训练提供准确的目标信息。然而，许多视觉跟踪方法忽略了尺度估计问题或采用启发式和穷举的尺度估计策略。在本文中，我们提出了一种新的基于相关滤波器的视觉跟踪方法，揭示了尺度估计和检测响应之间的缺失环节。与许多多尺度视觉跟踪器使用预先设计的标准生成不同尺度的样本，然后选择具有最大分类器响应的样本不同，本文推导了基于检测响应的尺度估计方程，从而可以数学地估计目标对象的尺度。为了获得更稳定的估计目标尺度，提出了一种考虑视觉跟踪先验知识的约束函数。此外，一个混合样本学习计划制定选择相关的训练样本具有较高的学习权重来训练的外观模型。我们的跟踪器的相关滤波器的框架下工作，以实现高跟踪速度。我们通过将我们提出的跟踪算法与其他14种最先进的跟踪器在对象跟踪基准（OTB）2013数据集中的所有视频序列上进行比较，证明了我们提出的跟踪算法的效率和鲁棒性。
In visual tracking, a mature scale estimation method can greatly improve tracking performance and provide accurate target information for model training. However, many visual tracking approaches ignore the scale estimation problem or adopt a heuristic and exhaustive scale-estimation strategy. In this paper, we propose a novel correlation-filter based visual tracking approach that reveals the missing link between scale estimation and the detection response. In contrast to many multi-scale visual trackers, which generate samples at different scales using some pre-designed criteria and then select the sample with the maximal classifier response, in this paper, we deduce a scale estimation equation based on detection responses; thus, the scale of the target object can be estimated mathematically. To obtain a more stable estimated object scale, a constraint function that considers the prior knowledge of visual tracking is proposed. Moreover, a hybrid sample learning scheme is formulated to select pertinent training samples with higher learning weights to train the appearance model. Our tracker operates under a framework of correlation filters to achieve a high tracking speed. We demonstrate the efficiency and robustness of our proposed tracking algorithm by comparing it with 14 other state-of-the-art trackers on all the video sequences in the object tracking benchmark (OTB) 2013 dataset.