Learn++ for Robust Object Tracking

Abstract

In this paper, a Learn++ (LPP) tracker is proposed to efficiently select specific classifiers for robust and long-term object tracking. In contrast to previous online methods, LPP tracker dynamically maintains a set of basic classifiers which are trained sequentially without accessing original data but preserving the previously acquired knowledge. The different subsets of basic classifiers can be specified to solve different sub-problems occurred in a non-stationary environment. Thus, an optimal classifier can be approximated in an active subspace spanned by selected adaptive basic classifiers. As a result, LPP tracker can address the "concept drift", by automatically adjusting the active subset and searching the optimal classifier in an active subspace spanned by the subset according to the distribution of the samples and recent performance. Experimental results show that LPP tracker yields state-of-the-art performance under various challenging environmental conditions and, especially, can overcome several challenges simultaneously.

Texture-Independent Long-Term Tracking Using Virtual Corners

“…Babenko et al [1] use multiple instance learning, instead of traditional supervised learning, for a more robust tracker. Zheng et al [32] address drift during tracking using a dynamic set of basis classifiers, employing different basis classifiers for different problems. All of these, however, suffer from the need to convert the estimated object position into a set of labelled training examples, and to couple the objective for the classifier (label prediction) to the objective for the tracker (object position estimation), which is difficult to perform optimally.…”