Adaptive cascaded and parallel feature fusion for visual object tracking

Due to its quick tracking, simple deployment, and straightforward principle, correlation filter-based tracking methods continue to have significant research implications. In order to make full use of different features while balancing the tracking speed and performance, the adaptive cascaded and parallel feature fusion-based tracker (ACPF), which could estimate the position, rotation and scale, respectively, is proposed. Comparing with other correlation filter-based trackers, the ACPF could fuse deep and handcrafted features in both Log-Polar and Cartesian branch and update templates according to the weights of response maps adaptively. Adaptive linear weights (ALW) are proposed to fuse feature response maps adaptively in the Log-Polar coordinates branch to improve the estimation of scale and rotation of object by solving constrained optimization problems. Response maps of shallow and deep features are merged adaptively by cascading numerous ALW modules in the Cartesian branch to make better utilize shallow and deep feature, and increase tracking accuracy. The final results are computed simultaneously by Cartesian and Log-Polar branch in parallel. Additionally, the learning rates are automatically changed in accordance with the weights of the ALW module to execute the adaptive template update. Extensive experiments on benchmarks show that the proposed tracker achieves the comparable results, especially in dealing with the challenges of deformation, rotation and scale variation.