Multi-Attention Fusion Lightweight Single Object Tracking Algorithm

With the continuous development of mobile application platforms, balancing real-time performance and accuracy has become a major challenge for object tracking technology, especially on resource-limited mobile devices. To address these challenges, a novel tracking algorithm called NanoMViT (Nano Mobile Vision Transformer for Object Tracking) and the ECSA (Efficient Channel and Spatial Attention) multi-attention module were proposed. By optimizing the backbone network architecture, the parameters of each module were significantly reduced. The final algorithm has a parameter count of only 709.134k and MACs of only 264.127M. Compared to the baseline algorithm, the parameter count is reduced by 66.3%, and the computational load is reduced by 50.2%. To validate the algorithm's performance, comprehensive evaluations were conducted on multiple standard tracking datasets, including VOT2018, DTB70, and LaSOT. The results show that the algorithm outperforms the baseline in aspects such as rapid deformation, tracking of similar objects, and performance under complex lighting conditions. These findings demonstrate that the algorithm has high practicality and efficiency in mobile tracking applications, but further validation is required in more complex environments.