Visual tracking via confidence template updating spatial-temporal regularized correlation filters

As an accurate and robust method for visual object tracking task, discriminative correlation filter (DCF) framework has received significant popularity from various researchers. For a given tracking sequence, the DCF formulates the target appearance model and tracking model according to the target information of the first frame, and then predicts the location and the scale of tracking target via specific tracking strategy while the tracking model and appearance model are updated via model learning strategy. However, the tracking performance of DCF tracker is limited by the undesirable impact of boundary effect caused by algorithm deficiency and response aberrance arisen from complex tracking environment. Aiming at tackling the above difficulties, a visual tracking method based on spatial-temporal regularized correlation filter with confidence template updating is developed, in which spatial-temporal regulariziers are formulated for tracking model learning process to tackle the spawning noises in the model learning process and an adaptive model template updating strategy for tracking strategy process is adopted to repress the response aberrance effect. The main findings and scientific contributions of our method (Ours) includes: 1) a novel spatial regularization method is introduced to restrain the boundary effect and to improve the overall tracking performance by penalizing the edge coefficient of correlation filter; 2) aiming at addressing the appearance variation of tracking target, a novel temporal regularizer is suggested to formulate a more stable learning process for the tracking model and further surmount the model noise caused by deficient model learning; 3) a novel adaptive updating strategy of model template is provided to alleviate the aberrances of response representations and obtain more accurate target prediction results. Extensive experimental results with 351 challenging videos on various datasets OTB2013, OTB2015, Temple-Color and UAV123 have proven that Ours can achieve favorable performances against other state-of-the-art trackers and efficiently adapt to a variety of complex scenarios in the tracking task.