Exploiting Hierarchical Response Fusion for Correlation Filter-Based UAV Tracking

The correlation filter (CF)-based methods have made significant progress in unmanned aerial vehicle (UAV) object tracking over the past decade. However, how to improve the model's discriminative ability and alleviate model drifts remain open issues for real-time UAV tracking. This paper tackles these challenges by developing a novel CF-based tracking algorithm with a hierarchical response fusion scheme and dynamic model update strategy. Specifically, we first perform intra-model response fusion for learning complementary features to train more robust correlation filters. Then, with the help of the Bayesian model, the discriminative ability of the CF model can be further enhanced via inter-model response fusion. Moreover, we introduce a simple yet effective model update strategy to prevent filter degradation and alleviate model drifts. Extensive experiments on two public UAV tracking benchmarks demonstrate the accuracy and robustness of the proposed tracker against other state-of-the-art trackers. With an average speed of 66 fps on a single CPU, the proposed algorithm has phenomenal practicability in real-time UAV applications.