SAM-Empowered Multidimensional Feature Fusion TIR Target Tracking Method for Internet of Things

Thermal Infrared (TIR) target tracking is widely used in smart cities and intelligent transportation systems within the Internet of Things (IoT) due to its insensitivity to lighting conditions. However, existing deep learning-based TIR tracking algorithms struggle with stable target tracking in complex backgrounds and interference due to limited training data and the unique properties of infrared images. To address these challenges, this paper introduces a Segment Anything Module (SAM)-Empowered Thermal Infrared Tracking Data Augmentation Method, which utilizes SAM’s strong segmentation capabilities to generate high-quality labeled data from existing datasets for training TIR trackers. We also propose a Multidimensional Perception Infrared Target Tracking Algorithm that enhances the integration of multi-level infrared features across spatial and channel dimensions, improving feature representational strength and tracking robustness. Comparative experiments on the LSOTB-TIR dataset show our method achieving 82.9% Precision and 74.3% Normalized Precision, surpassing the current leading algorithm by 0.4% and 0.2%, respectively, while also achieving a speed improvement of 9 fps. Similarly, on the PTB-TIR dataset, our method attains a top Success score of 68.5%, leading the state-of-the-art by 0.3% in Success and 39.4 fps in tracking speed, showcasing its effectiveness and efficiency. • Propose a Segment Anything Model-based TIR tracking data augmentation method. • A TIR target tracking method based on multi-level feature fusion is proposed. • Our method achieves excellent performance on two commonly used TIR tracking benchmarks.