SCSTCF: Spatial-Channel Selection and Temporal Regularized Correlation Filters for visual tracking

Recently, combining multiple features into discriminative correlation filters to improve tracking representation has shown great potential in object tracking. Existing trackers apply fixed weights to fuse features or fuse response maps, which cannot adapt to the object drift well. Moreover, in the tracking algorithm, using cyclic shift to obtain training samples always cause boundary effect, resulting in dissatisfied tracking effect. Therefore, we first design a multiple features fusion method. Various handcrafted features are fused with the same weight, then the fused handcrafted features and deep features are fused by adaptive weights, which considerably improves the representation ability of the tracking object. Second, we propose a correlation filter object function model called Spatial-Channel Selection and Temporal Regularized Correlation Filters. We perform the grouping features selection from the dimensions of channel, spatial and temporal, so as to establish the relevance between the multi-channel features and the correlation filter. Finally, we transform the objective function of the model with equality constraint to augmented Lagrangian multiplier formula without constraint, which is divided into three subproblems with closed-form solutions. The optimal solution is obtained by iteratively solving three subproblems using Alternating Direction Multiplier Method (ADMM). We conduct extensive experiments in four public datasets, OTB-2013, OTB-2015, TC128, UAV123, and VOT2016. The experimental results represent our proposed tracker performs favorably against other prevailing trackers in success rate and precision. • We propose an adaptive weight fusion method to fuse handcrafted features and deep feature response maps. • We propose a novel CF model which combine spatial-channel selection of feature maps with temporal consistency constraint. • Our model is a general CF model and is derived by ADMM to obtain its optimal closed-form solution. • We achieve comparable performances with other state-of-the-art methods on 5 challenging datasets.