Rotation-aware dynamic temporal consistency with spatial sparsity correlation tracking

Abstract

Recently, discriminative correlation filter (DCF)-based trackers have been widely applied to visual tracking. However, a significant problem of DCF-based trackers is that the model uses the fixed patterns of temporal modeling and fails to suppress the distractive features. To obviate the issue, we propose the dynamic temporal consistency with spatial sparsity correlation filter. The dynamics refers to the adaptive temporal consistency hidden in the response maps and dynamic lasso constraint moderated by the prior knowledge. Unlike the classical temporal modeling method applied to filter model, we exploit the consistency of the response maps to perceive adaptive temporal continuity modeling to enable the filter to have self-regularized ability. Temporal modeling and spatial sparsity are incorporated in a unified optimization learning model and optimized together with ADMM algorithm and Sherman-Morrison formula. In particular, the tracking performance is hindered by the weak capacity to rotation variation. For the sake of estimating the rotation angle accurately to maintain rotation invariance, we explore the coarse-to-fine rotation estimation module. The coarse rotation is supported by the angle pool and the part-based tracker. By introducing the connected hyper ellipse fitting strategy to eliminate fake distractors to ensure a pure target region, the fine level attempts to achieve the optimal rotation angle with the minimum second order statistical bias. The design enables the filter to train with the recovered rotated image instead of axis-aligned bounding box, which attributes to alleviating the impact of ambiguous region and concentrating on the interested target. Extensive experimental results validate the superiority of the proposed method against other state-of-the art trackers and exhibit a remarkable generality in the rotated challenging scenarios. • Dynamic temporal modeling can make the filter highlight the ever-changing region. • Dynamic temporal consistency modeling and spatial sparsity are incorporated in a unified optimization learning model. • The connected hyper ellipse fitting strategy tries to achieve the optimal rotation angle. • Comprehensive experiments have fully demonstrated the superior performance of our proposed method.