Adaptive Gaussian-Like Response Correlation Filter for UAV Tracking

Existing Discriminative Correlation Filters (DCF) based methods usually use a heatmap with a two-dimensional Gaussian distribution to represent the foreground probability map on the search image plane and regularize tracker’s response map in every frame. However, in many real-world scenarios, there often exist non-zero correlation responses in background regions due to the existence of distractors that have similar appearance as the target. In such cases, forcing the output response map to be ideally Gaussian-distributed will lead to contradictory constrains on representation learning, thus hurting performance. To alleviate this, we propose a novel tracker, named Gaussian-like response Correlation Filter (GLCF), which constructs expected response maps by assigning non-zero values to the locations of distractors adaptively depending on their similarity with the target, and thus maintains inter-object consistency and improves robustness. Extensive experiments on four benchmarks well demonstrate the superiority of the proposed method over both DCF and deep based trackers. Specifically, our method achieves new state-of-the-art performance on UAVDT dataset while running at a speed of 28.9 FPS on a single CPU.