A Meta-Q-Learning Approach to Discriminative Correlation Filter based Visual Tracking
A Meta-Q-Learning Approach to Discriminative Correlation Filter based Visual Tracking
复制标题
DOI：
10.1007/s10846-020-01273-2
复制
发表时间：
2020-12
影响因子：
3.3
通讯作者：
Akihiro Kubo;Kourosh Meshgi;S. Ishii
中科院分区：
文献类型：
--
作者：
Akihiro Kubo;Kourosh Meshgi;S. Ishii
视觉目标跟踪仍然是一个具有挑战性的计算机视觉问题与许多现实世界的应用。基于鉴别相关滤波器（DCF）的方法是最近处理这个问题的最先进的方法。无论情况如何，应用DCF时的学习率通常是固定的。然而，这个速率对于鲁棒跟踪是重要的，因为现实世界的视频序列包括各种动态变化，诸如遮挡、运动模糊和变形。在这项研究中，我们提出了元Q学习相关滤波器（MQCF），一种基于手工制作的方向梯度直方图（HOG）特征，通过强化学习动态确定基于基线DCF的跟踪器的学习率的方法。强化学习的结合使我们能够为图像补丁训练一个函数，该函数以自主的方式输出基线跟踪器的依赖于情况的学习率。我们使用两个开放基准测试（即OTB-2015和VOT-2105）评估了这种方法，发现我们的MQCF跟踪器在OTB-2015上的曲线下面积方面优于基线最先进的跟踪器1.8%，在VOT-2015挑战中的预期平均重叠方面相对增益为8.4%。我们的研究结果证明了所谓的元学习与基于DCF的视觉对象跟踪的优势。
Visual object tracking remains a challenging computer vision problem with numerous real-world applications. Discriminative correlation filter (DCF)-based methods are a recent state-of-the-art approach for dealing with this problem. The learning rate when applying a DCF is typically fixed, regardless of the situation. However, this rate is important for robust tracking, insofar as real-world video sequences include a variety of dynamical changes, such as occlusions, motion blur, and deformations. In this study, we propose Meta-Q-learning Correlation Filter (MQCF), a method for dynamically determining the learning rate of a baseline DCF-based tracker based on hand-crafted features of Histogram of Oriented Gradient (HOG), by means of reinforcement learning. The incorporation of reinforcement learning enables us to train a function for an image patch that outputs a situation-dependent learning rate of the baseline tracker in an autonomous fashion. We evaluated this method using two open benchmarks, namely, OTB-2015 and VOT-2105, and found our MQCF tracker outperformed a baseline state-of-the-art tracker by 1.8% in Area Under Curve on OTB-2015, and 8.4% relative gain in Expected Average Overlap in the VOT-2015 challenge. Our results demonstrate the advantages of the so-called meta-learning with DCF-based visual object tracking.