Learning region sparse constraint correlation filter for tracking

Correlation filters (CF) have sparked a lot of interest in visual tracking owing to their impressive performance. Most of previous correlation filter methods learn a filter from all features of the sample region. However, some features may be distractive, like those from occlusions and deformations, leading to tracking drift in the next frame. To mitigate this problem, we propose a novel region sparse constraint correlation filter (RSCF) to adaptively ignore those distractive features. The proposed method is formulated based on the elastic net model, and a binary mask is used to directly limit the sparsity of the filter values corresponding to the target region instead of the whole sample region. Besides, a context-aware term is integrated to enhance the discriminant ability of the filter. Last, an ADMM optimization algorithm is proposed to solve the model. Qualitative evaluations have been conducted on well-known benchmark, such as OTB-2013, OTB-2015, Temple-Color 128 and VOT2016. Experiment results demonstrate that the proposed tracker performs favorably against several state-of-the-art methods.