Visual Tracking Using Attention-Modulated Disintegration and Integration

题目：Visual Tracking Using Attention-Modulated Disintegration and Integration

来源：CVPR 2016

Abstract

本文使用注意力调制的解体和整合进行跟踪，把一个目标分解成多个认知单元，并且训练多个基础的跟踪器，目的是根据不同的特征和核的类型来调制注意力的分布。在整合阶段，重新把这些单元连接起来，以有效的记忆和识别目标物体。关于基础的跟踪器，作者提出了一种新颖的基于注意力特征的相关滤波器（AtCF）focuses on distinctive attentional features.

1.Introduction

本文提出使用一种结构化的认知模型来做目标跟踪（SCT）。该SCT由两个独立阶段组成：解体（disintegration）和整合（integration）。在解体阶段，目标被分成许多小的认知结构单元，which are memorized separately。每个单元包括a specific color or a distinguishable target shape，也就是说每个单元要么包括颜色，要么包括形状，二选一，这个是作为特征的。然后每个单元用拥有不同核类型的基础跟踪器进行训练。

在整合阶段，结构单元进行充分的连接和记忆，以表达目标的外观。当遇到目标的外观变化的时候，SCT就运用来自认知单元（memorized in the disintegration stage）的所有响应，然后通过更好的对认知单元进行连接来识别目标，referring to the memorized combinations.

通过这两大阶段，SCT就能够对于复杂的输入产生非常鲁棒的响应了。当背景杂波出现时，disintegration阶段就训练这些认知单元，以提升判别力。此外，还要学习这些认知单元的最好的结合，并且在integration阶段进行存储。如果目标的外观突然发生变化的话，the integration 就降低the influence of the previously influential cognitive units and enlarges the influence of new appropriate units,这会帮助SCT可以适应目标的快速变化。

关于基础滤波器（elementary trackers），作者提出一种基于注意力特征的相关滤波器AtCF。这个AtCF集中在the attentional features discriminated from the background,这点是深受人类的认知模型启发的。每个AtCF由一个注意力权重估计器（an attentional weight  estimator）和一个核化的相关滤波器（KCF）组成。As a result of AtCF， 本文所提的SCF方法可以避免the problems caused by background features misplaced in the bounding box, such as drift and a missing target. In the disintegration stage, multiple AtCFs are updated using various feature and kernel types. The integration stage combines the responses of AtCFs by ordering the AtCFs following their performance.也就是说，本文方法可以应对drift和missing target的问题（不过好像每篇文章都会这样说哦）。

2.Related Research

相关滤波法很流行啊！不过呢，之前的相关滤波法使用的是一种固定类型的特征和核，所以在性能上是有所局限的。在本文中，为增强目标的判别力，multiple correlation filters using various types of feature and kernel are updated in the disintegration stage。当然了，已经有一些将tracker进行combile以生成一个strong tracker的方法，如boosting和bootstrap learning。Unlike the trackers based on boosting and bootstrap learning, elementary trackers of the proposed algorithm are updated independently using samples only from the current frame.The independent update enables the use of a large number of negative samples to improve the performance. Contrary to the previous trackers based on the multiple weak trackers, the proposed algorithm uses a fast correlation filter as weak trackers to work in real-time. In addition, against the hybrid trackers fusing the  trackers of different structure, the proposed framework can be extended easily by supplementing additional feature or kernel type.

在saliency map的使用方面，the proposed framework does not find the target location directly from saliency map but use the saliency map as a weight map for correlation filters. In addition, the attentional scheme of the proposed algorithm  includes both top-down (in the tracking phase) and bottom-up (in the updating phase) factors.

3.Proposed Tracker

In the disintegration stage, multiple AtCFs are generated and updated to cover the various properties of the target.每个AtCF使用各自特有的特征和核的组合。～～

略

4.Experimental  Result

略

5.Conclusion

In the disintegration stage, the target was trainedby multiple elementary trackers with various types of feature and kernel, which improved the ability to discriminate the target from background clutter.

In the integration stage, the responses of the multiple elementary trackers were combined according to the memorized priorities and reliabilities, with quick adaptation against sudden changes of the target’s appearance.

As the elementary tracker, the AtCF was suggested, which demonstrated robustness to partial occlusion and drift by reducing the weights for the background features in the bounding box of the target.