SPECTER: A Tracker with Super-Resolution Siamese Network Reconstruction and Multi-Scale Feature Extraction for Thermal Infrared Pedestrian Tracking
Ruoyan Xiong, Shang Zhang, Yue Zhang, Ming Zhang · 2024
Existing thermal infrared (TIR) trackers attempt to acquire robust discriminative features of the target, as TIR tracking traditionally relies on learning an appearance model of the target. However, the challenge of low-resolution and sparse details in TIR images limits the effectiveness of existing trackers. To tackle this limitation, we introduce a novel tracker for robust TIR pedestrian tracking. Our proposed tracker emphasizes enhancing target discriminability through multi-scale deep feature extraction and high-resolution reconstruction. Firstly, the multi-scale feature extraction and fusion network captures diverse features using a multi-branch structure and incorporates grouped convolutions to extract varied deep features by increasing the cardinality of each group. Secondly, a super-resolution Siamese network is designed with a chain-like structure of stacked feature extraction functions. The core of this network is a Siamese residual block, which progressively aggregates image features through residual learning, extracting deep abstract feature information and improving model generalization. Moreover, by utilizing sub-pixel convolutional layers to reconstruct deep features, high-resolution TIR images are generated to improve target recognition. To the best of our knowledge, we are the first to perform super-resolution technology for TIR target tracking. Extensive experiment results on four TIR tracking benchmarks including LSOTB-TIR, PTB-TIR, VOT-TIR2015, and VOT-TIR 2017 demonstrate that our tracker achieves the best performance compared to other state-of-the-art trackers.