Aggregating Correlation Filter with Multiple Kernels Learning for Robust Visual Object Tracking.

Yanxia Wei,Hairong Xiao,Deying Feng,Jian Mu,Jie Yang

DOI: https://doi.org/10.1145/3532213.3532305

2022-01-01

Abstract:Recently, Discriminative Correlation Filter based trackers have increasingly become popular in the domain of visual object tracking, which is benefited by their effective and robustness in terms of tracking performance. However, the significant variations of target appearance, such as Occlusion or Deformation, always suppress the performance improvement. Although the multi-feature fusion strategy partly alleviates the above factors, the synthetic scheme also ambiguates the overall presentation ability of multiple features and results in degradation of discriminative power during classification filter learning. In this paper, we present a novel tracking framework based on the correlation filter with multiple kernel learning. Instead of directly integrating different features, we decompose a tracking process into several module, including low-dimension modules and high-dimension modules. The handcraft features extracted from low-dimension modules are robust to the small variation of target while the semantic convolutional features from the high-dimension modules can distinguish the target object from the relative complex tracking environment. In addition, we also design a reliability detection mechanism based on the confidence map to accurately measure the important degree of each kernel learning and precisely estimate the target state in the subsequent frames. For verifying our proposed tracker, we conduct extensive experiments on several standard sequence benchmark. Experiment results show that our proposed trackers perform favorably against other state-of-the-art trackers.