SRCFT: A Correlation Filter Tracker with Siamese Super-Resolution Network and Sample Reliability Awareness for Thermal Infrared Target Tracking

Correlation filter (CF) trackers have received significant attention for their computational efficiency in thermal infrared (TIR) target tracking. However, the challenge of low-resolution TIR images and the presence of unreliable samples during CF training process limit the effectiveness of existing trackers. To address these issues, we introduce a novel CF tracker called SRCFT, which incorporates a Siamese super-resolution network and sample reliability awareness. Firstly, we refine the objective function by introducing sparse regularization and spatial regularization strategies during the training process of SRCFT, with the aim of improving spatial consistency and model stability. Secondly, we propose a Siamese super-resolution network that seamlessly incorporates continuous functions and kernel grid iterations into the tracker learning process. This network is designed to generate high-resolution images of various scales from low-resolution TIR images. Thirdly, we propose a sample reliability awareness strategy to mitigate the contamination of negative sample during training. This strategy dynamically adjusts the update formula based on the perceived reliability of samples, reducing the influence of negative samples through adjusted weighting. To the best of our knowledge, we are the first to perform sample reliability awareness and super-resolution technology for TIR target tracking. Extensive experiments are conducted on four TIR tracking benchmarks, including LSOTB-TIR, PTB-TIR, VOT-TIR2015, and VOT-TIR 2017. The relevant experimental results clearly demonstrate that our SRCFT outperforms the state-of-the-art trackers.