Collaborative Correlation Tracking

Motivation. Correlation filter based tracking has attracted many researchers’ attention in recent years for high efficiency and robustness. Most existing works [1, 2, 4] focus on exploiting different characteristics with correlation filters for visual tracking, e.g., circulant structure, kernel trick, effective feature representation and context information. Despite its good performance, most of these correlation methods have two main limitations, the first is how to adjust the object scale efficiently. In order to consistently track the object, Danelljan et.al. [2] proposed a separate 1-dimensional correlation filter to estimate the target scale, but they only use the original feature space as the object representation. The second limitation is how to handle the model drift problem caused by the long-term occlusion or out-of-view, which is a very important problem for online tracking [5]. One common mechanism is to introduce a detection module which can select some effective candidates to rectify the base correlation tracker. Contributions. In this paper, we design a novel online CUR filter for detection. CUR matrix approximation computes the low rank approximation of a given matrix by using the actual rows and columns of the matrix and have been studied in the area of theoretical computer science for large matrix approximation [3]. In the long-term tracking process, all of the historical object representations can form a large data matrix for the current frame which fits for the CUR theory. The large data matrix can be fast approximated by online CUR for representing the intrinsic object structure. In this work we develop an online CUR for learning an online detection filter by random sampling. The online CUR filter can not only exploit the low rank property of object representation [7] in the spatial-temporal domain of tracking, but also project the historical object representation matrix into a subspace with error upper bound so as to achieve a robust object representation. The low rank property of object representation is prevalent in long-term tracking and could be used to alleviate the model drift. Meanwhile, we propose multi-scale kernelized correlation filter as our tracking filter by embedding the scale variation into the kernelized correlation filter while forming a separate pyramid of object representation. The collaboration between CUR filter and multi-scale kernelized correlation filter constructs the proposed tracker–Collaborative Correlation Tracker (CCT)1.