Global Context-Aware Single Object Tracking Algorithm Based on Enhanced Swin-T

Single-object tracking based on computer vision facilitates the collection, analysis, and utilization of target information. Currently, the latest single-object tracking algorithms solely rely on the Transformer for feature fusion and matching between the template frame and the search area frame. However, all of the Transformer-based algorithms fail to fully leverage the dynamic encoding capabilities of the Transformer for feature extraction, neglecting the inherent abilities of the Vision Transformer in feature extraction and decoding prediction. To address these issues, we propose a global context-based single-object tracking method with an enhanced tiny Swin Transformer encoder (Swin-T). By leveraging the Swin-T as our backbone network, we are able to extract both local and global information through the use of shifting windows. This approach allows for a more precise representation of the object state, consequently enhancing the accuracy of a single-object tracking system. Through comparative experiments on multiple challenging datasets, our algorithm demonstrates an increase in success rates of 1.9% and 2.8% on the Lasot and UAV123 datasets, respectively, and an improvement of 0.5% in the F-score on the VOT2020LT dataset compared to the benchmark algorithm.