Adaptive Low-rank Subspace Learning With Online Optimization For Robust Visual Tracking

As one of the fundamental problems of computer vision, visual tracking has numerous advanced applications such as video surveillance, medical image analysis, gesture recognition and human computer interaction. In the past decades, although various algorithms have been proposed for visual, visual tracking is an challenging task due to heavy occlusion, illumination change, background clutter and complex object motion happen in complex and dynamic scenes.Subspace representation is robust for the appearance changes caused by external factors and internal motion and easy to compute. So various subspace based algorithms have been ap-plied to visual tracking. However, these methods only consider the sparsity or low-rankness of the coefficients, which is not sufficient enough for appearance subspace learning on complex video sequences. Moreover, as both the low-rank and the column sparse measures, it is chal-lenging to incrementally solve optimization problems with both nuclear norm and column sparse norm. Based on the above discussions, the paper proposes a novel low-rank subspace learning with adaptive penalization model to learn low-rank features of the target and designs an online optimization algorithm to apply the model to visual tracking.First, different from previous work, which often simply decompose observations as low-rank features and sparse errors the proposed algorithm can simultaneously learns the sub-space basis, low-rank coefficients and column sparse errors to formulate appearance subspace. Second, we introduce a Hadamard production based regularization to incorporate rich gener-ative/discriminative structure constraints to adaptively penalize the coefficients for subspace learning. It is shown that such adaptive penalization can significantly improve the robustness on severely corrupted data set. Finally, to utilize our algorithm for online visual tracking, we also develop an efficient incremental optimization scheme for nuclear norm and column sparse norm minimizations. Experiments on 50 challenging video sequence demonstrate that our algo-rithm is robust and stability, and it can handle sequence sets with different properties at the same time and outperform other state-of-the-art methods.