Adaptive Correlation Filters Feature Fusion Learning for Visual Tracking

Tracking algorithms based on discriminative correlation filters (DCFs) usually employ fixed weights to integrate feature response maps from multiple templates. However, they fail to exploit the complementarity of multi-feature. These features are against tracking challenges, e.g., deformation, illumination variation, and occlusion. In this work, we propose a novel adaptive feature fusion learning DCFs-based tracker (AFLCF). Specifically, AFLCF can learn the optimal fusion weights for handcrafted and deep feature responses online. The fused response map owns the complementary advantages of multiple features, obtaining a robust object representation. Furthermore, the adaptive temporal smoothing penalty adapts to the tracking scenarios with motion variation, avoiding model corruption and ensuring reliable model updates. Extensive experiments on five challenging visual tracking benchmarks demonstrate the superiority of AFLCF over other state-of-the-art methods. For example, AFLCF achieves a gain of 1.9\(\%\) and 4.4\(\%\) AUC score on LaSOT compared to ECO and STRCF, respectively.